renderdiff: Use single lib for osmesa (#8255)

Previously, we linked against libOSMesa through the linker and
then used dlopen to link against libGL. This is not how libOSMesa
is intended to be used.  Instead, we use dlopen on libOSMesa
(via bluegl), which then will map the correct GL methods for us,
and for the OSMesa functions, we also map those functions
from libOSMesa (instead of relying on compile-time linker).

.github/actions/android-continuous/action.yml

@@ -0,0 +1,17 @@
+name: 'Android Continuous'
+inputs:
+  build-abi:
+    description: 'The target platform ABI'
+    required: true
+    default: 'armeabi-v7a'
+runs:
+  using: "composite"
+  steps:
+    - uses: actions/setup-java@v3
+      with:
+        distribution: 'temurin'
+        java-version: '17'
+    - name: Run build script
+      run: |
+        cd build/android && printf "y" | ./build.sh continuous ${{ inputs.build-abi }}
+      shell: bash

.github/actions/ubuntu-apt-add-src/action.yml

@@ -0,0 +1,9 @@
+name: 'ubuntu apt add deb-src'
+runs:
+  using: "composite"
+  steps:
+    - name: "ubuntu apt add deb-src"
+      run: |
+          echo "deb-src http://archive.ubuntu.com/ubuntu jammy main restricted universe" | sudo tee /etc/apt/sources.list.d/my.list
+          sudo apt-get update
+      shell: bash

.github/workflows/android-continuous.yml

@@ -10,30 +10,13 @@ on:
 jobs:
   build-android:
     name: build-android
-    runs-on: macos-14
+    # We intentially use a larger runner here to enable larger disk space
+    # (standard linux runner will fail on disk space and faster build time).
+    runs-on: ubuntu-22.04-32core
 
     steps:
-      - uses: actions/checkout@v3.3.0
-      - uses: actions/setup-java@v3
+      - uses: actions/checkout@v4.1.6
+      - name: Run Android Continuous
+        uses: ./.github/actions/android-continuous
         with:
-          distribution: 'temurin'
-          java-version: '17'
-      - name: Run build script
-        run: |
-          cd build/android && printf "y" | ./build.sh continuous
-      - uses: actions/upload-artifact@v1.0.0
-        with:
-          name: filament-android
-          path: out/filament-android-release.aar
-      - uses: actions/upload-artifact@v1.0.0
-        with:
-          name: filamat-android-full
-          path: out/filamat-android-release.aar
-      - uses: actions/upload-artifact@v1.0.0
-        with:
-          name: gltfio-android-release
-          path: out/gltfio-android-release.aar
-      - uses: actions/upload-artifact@v1.0.0
-        with:
-          name: filament-utils-android-release
-          path: out/filament-utils-android-release.aar
+          build-abi: armeabi-v7a,arm64-v8a,x86_64

.github/workflows/ios-continuous.yml

@@ -10,14 +10,14 @@ on:
 jobs:
   build-ios:
     name: build-ios
-    runs-on: macos-14
+    runs-on: macos-14-xlarge
 
     steps:
-      - uses: actions/checkout@v3.3.0
+      - uses: actions/checkout@v4.1.6
       - name: Run build script
         run: |
           cd build/ios && printf "y" | ./build.sh continuous
-      - uses: actions/upload-artifact@v1.0.0
+      - uses: actions/upload-artifact@v4
         with:
           name: filament-ios
           path: out/filament-release-ios.tgz

.github/workflows/linux-continuous.yml

@@ -10,14 +10,14 @@ on:
 jobs:
   build-linux:
     name: build-linux
-    runs-on: ubuntu-22.04
+    runs-on: ubuntu-22.04-16core
 
     steps:
-      - uses: actions/checkout@v3.3.0
+      - uses: actions/checkout@v4.1.6
       - name: Run build script
         run: |
           cd build/linux && printf "y" | ./build.sh continuous
-      - uses: actions/upload-artifact@v1.0.0
+      - uses: actions/upload-artifact@v4
         with:
           name: filament-linux
           path: out/filament-release-linux.tgz

.github/workflows/mac-continuous.yml

@@ -10,14 +10,14 @@ on:
 jobs:
   build-mac:
     name: build-mac
-    runs-on: macos-14
+    runs-on: macos-14-xlarge
 
     steps:
-      - uses: actions/checkout@v3.3.0
+      - uses: actions/checkout@v4.1.6
       - name: Run build script
         run: |
           cd build/mac && printf "y" | ./build.sh continuous
-      - uses: actions/upload-artifact@v1.0.0
+      - uses: actions/upload-artifact@v4
         with:
           name: filament-mac
           path: out/filament-release-darwin.tgz

.github/workflows/npm-deploy.yml

@@ -13,7 +13,7 @@ jobs:
     name: npm-deploy
     runs-on: macos-14
     steps:
-      - uses: actions/checkout@v3.3.0
+      - uses: actions/checkout@v4.1.6
         with:
           ref: ${{ github.event.inputs.release_tag }}
       # Setup .npmrc file to publish to npm

.github/workflows/presubmit.yml

@@ -15,10 +15,10 @@ jobs:
 
     strategy:
       matrix:
-        os: [macos-14, ubuntu-22.04]
+        os: [macos-14-xlarge, ubuntu-22.04-16core]
 
     steps:
-      - uses: actions/checkout@v3.3.0
+      - uses: actions/checkout@v4.1.6
       - name: Run build script
         run: |
           WORKFLOW_OS=`echo \`uname\` | sed "s/Darwin/mac/" | tr [:upper:] [:lower:]`
@@ -29,10 +29,10 @@ jobs:
 
   build-windows:
     name: build-windows
-    runs-on: windows-2019
+    runs-on: win-2019-16core
 
     steps:
-      - uses: actions/checkout@v3.3.0
+      - uses: actions/checkout@v4.1.6
       - name: Run build script
         run: |
           build\windows\build-github.bat presubmit
@@ -40,24 +40,26 @@ jobs:
 
   build-android:
     name: build-android
-    runs-on: macos-14
+    runs-on: ubuntu-22.04-16core
 
     steps:
-      - uses: actions/checkout@v3.3.0
+      - uses: actions/checkout@v4.1.6
       - uses: actions/setup-java@v3
         with:
           distribution: 'temurin'
           java-version: '17'
       - name: Run build script
+        # Only build 1 64 bit target during presubmit to cut down build times during presubmit
+        # Continuous builds will build everything
         run: |
-          cd build/android && printf "y" | ./build.sh presubmit
+          cd build/android && printf "y" | ./build.sh presubmit arm64-v8a
 
   build-ios:
     name: build-iOS
-    runs-on: macos-14
+    runs-on: macos-14-xlarge
 
     steps:
-      - uses: actions/checkout@v3.3.0
+      - uses: actions/checkout@v4.1.6
       - name: Run build script
         run: |
           cd build/ios && printf "y" | ./build.sh presubmit
@@ -67,10 +69,25 @@ jobs:
 
   build-web:
     name: build-web
-    runs-on: macos-14
+    runs-on: ubuntu-22.04-16core
 
     steps:
-      - uses: actions/checkout@v3.3.0
+      - uses: actions/checkout@v4.1.6
       - name: Run build script
         run: |
           cd build/web && printf "y" | ./build.sh presubmit
+
+  test-renderdiff:
+    name: test-renderdiff
+    runs-on: ubuntu-22.04-32core
+
+    steps:
+      - uses: actions/checkout@v4.1.6
+      - uses: ./.github/actions/ubuntu-apt-add-src
+      - name: Run script
+        run: |
+          source ./build/linux/ci-common.sh && bash test/renderdiff_tests.sh
+      - uses: actions/upload-artifact@v4
+        with:
+          name: presubmit-renderdiff-result
+          path: ./out/renderdiff_tests

.github/workflows/release.yml

@@ -31,7 +31,7 @@ jobs:
 
     strategy:
       matrix:
-        os: [macos-14, ubuntu-22.04]
+        os: [macos-14-xlarge, ubuntu-22.04-32core]
 
     steps:
       - name: Decide Git ref
@@ -41,7 +41,7 @@ jobs:
           TAG=${REF##*/}
           echo "ref=${REF}" >> $GITHUB_OUTPUT
           echo "tag=${TAG}" >> $GITHUB_OUTPUT
-      - uses: actions/checkout@v3.3.0
+      - uses: actions/checkout@v4.1.6
         with:
           ref: ${{ steps.git_ref.outputs.ref }}
       - name: Run build script
@@ -65,7 +65,7 @@ jobs:
 
   build-web:
     name: build-web
-    runs-on: macos-14
+    runs-on: ubuntu-22.04-16core
     if: github.event_name == 'release' || github.event.inputs.platform == 'web'
 
     steps:
@@ -76,7 +76,7 @@ jobs:
           TAG=${REF##*/}
           echo "ref=${REF}" >> $GITHUB_OUTPUT
           echo "tag=${TAG}" >> $GITHUB_OUTPUT
-      - uses: actions/checkout@v3.3.0
+      - uses: actions/checkout@v4.1.6
         with:
           ref: ${{ steps.git_ref.outputs.ref }}
       - name: Run build script
@@ -98,7 +98,7 @@ jobs:
 
   build-android:
     name: build-android
-    runs-on: macos-14
+    runs-on: ubuntu-22.04-16core
     if: github.event_name == 'release' || github.event.inputs.platform == 'android'
 
     steps:
@@ -109,7 +109,7 @@ jobs:
           TAG=${REF##*/}
           echo "ref=${REF}" >> $GITHUB_OUTPUT
           echo "tag=${TAG}" >> $GITHUB_OUTPUT
-      - uses: actions/checkout@v3.3.0
+      - uses: actions/checkout@v4.1.6
         with:
           ref: ${{ steps.git_ref.outputs.ref }}
       - uses: actions/setup-java@v3
@@ -120,7 +120,7 @@ jobs:
         env:
           TAG: ${{ steps.git_ref.outputs.tag }}
         run: |
-          cd build/android && printf "y" | ./build.sh release
+          cd build/android && printf "y" | ./build.sh release armeabi-v7a,arm64-v8a,x86,x86_64
           cd ../..
           mv out/filament-android-release.aar out/filament-${TAG}-android.aar
           mv out/filamat-android-release.aar out/filamat-${TAG}-android.aar
@@ -152,7 +152,7 @@ jobs:
 
   build-ios:
     name: build-ios
-    runs-on: macos-14
+    runs-on: macos-14-xlarge
     if: github.event_name == 'release' || github.event.inputs.platform == 'ios'
 
     steps:
@@ -163,7 +163,7 @@ jobs:
           TAG=${REF##*/}
           echo "ref=${REF}" >> $GITHUB_OUTPUT
           echo "tag=${TAG}" >> $GITHUB_OUTPUT
-      - uses: actions/checkout@v3.3.0
+      - uses: actions/checkout@v4.1.6
         with:
           ref: ${{ steps.git_ref.outputs.ref }}
       - name: Run build script
@@ -185,7 +185,7 @@ jobs:
 
   build-windows:
     name: build-windows
-    runs-on: windows-2019
+    runs-on: windows-2019-32core
     if: github.event_name == 'release' || github.event.inputs.platform == 'windows'
 
     steps:
@@ -197,7 +197,7 @@ jobs:
           echo "ref=${REF}" >> $GITHUB_OUTPUT
           echo "tag=${TAG}" >> $GITHUB_OUTPUT
         shell: bash
-      - uses: actions/checkout@v3.3.0
+      - uses: actions/checkout@v4.1.6
         with:
           ref: ${{ steps.git_ref.outputs.ref }}
       - name: Run build script
@@ -205,6 +205,7 @@ jobs:
           TAG: ${{ steps.git_ref.outputs.tag }}
         run: |
           build\windows\build-github.bat release
+          echo on
           move out\filament-windows.tgz out\filament-%TAG%-windows.tgz
         shell: cmd
       - uses: actions/github-script@v6

.github/workflows/web-continuous.yml

@@ -10,14 +10,14 @@ on:
 jobs:
   build-web:
     name: build-web
-    runs-on: macos-14
+    runs-on: ubuntu-22.04-16core
 
     steps:
-      - uses: actions/checkout@v3.3.0
+      - uses: actions/checkout@v4.1.6
       - name: Run build script
         run: |
           cd build/web && printf "y" | ./build.sh continuous
-      - uses: actions/upload-artifact@v1.0.0
+      - uses: actions/upload-artifact@v4
         with:
           name: filament-web
           path: out/filament-release-web.tgz

.github/workflows/windows-continuous.yml

@@ -10,15 +10,15 @@ on:
 jobs:
   build-windows:
     name: build-windows
-    runs-on: windows-2019
+    runs-on: windows-2019-32core
 
     steps:
-      - uses: actions/checkout@v3.3.0
+      - uses: actions/checkout@v4.1.6
       - name: Run build script
         run: |
           build\windows\build-github.bat continuous
         shell: cmd
-      - uses: actions/upload-artifact@v1.0.0
+      - uses: actions/upload-artifact@v4
         with:
           name: filament-windows
           path: out/filament-windows.tgz

CMakeLists.txt

@@ -45,6 +45,8 @@ option(FILAMENT_ENABLE_FEATURE_LEVEL_0 "Enable Feature Level 0" ON)
 
 option(FILAMENT_ENABLE_MULTIVIEW "Enable multiview for Filament" OFF)
 
+option(FILAMENT_SUPPORTS_OSMESA "Enable OSMesa (headless GL context) for Filament" OFF)
+
 set(FILAMENT_NDK_VERSION "" CACHE STRING
     "Android NDK version or version prefix to be used when building for Android."
 )
@@ -73,6 +75,10 @@ set(FILAMENT_BACKEND_DEBUG_FLAG "" CACHE STRING
     "A debug flag meant for enabling/disabling backend debugging paths"
 )
 
+set(FILAMENT_OSMESA_PATH "" CACHE STRING
+    "Path to the OSMesa header and lib"
+)
+
 # Enable exceptions by default in spirv-cross.
 set(SPIRV_CROSS_EXCEPTIONS_TO_ASSERTIONS OFF)
 
@@ -132,12 +138,22 @@ else()
 endif()
 
 if (LINUX)
+    if (NOT FILAMENT_OSMESA_PATH STREQUAL "")
+        if (NOT EXISTS ${FILAMENT_OSMESA_PATH}/)
+            message(FATAL_ERROR "Cannot find specified OSMesa build directory: ${FILAMENT_OSMESA_PATH}")
+        endif()
+        set(FILAMENT_SUPPORTS_OSMESA TRUE)
+    endif()
+
     if (FILAMENT_SUPPORTS_WAYLAND)
         add_definitions(-DFILAMENT_SUPPORTS_WAYLAND)
         set(FILAMENT_SUPPORTS_X11 FALSE)
     elseif (FILAMENT_SUPPORTS_EGL_ON_LINUX)
         add_definitions(-DFILAMENT_SUPPORTS_EGL_ON_LINUX)
         set(FILAMENT_SUPPORTS_X11 FALSE)
+    elseif (FILAMENT_SUPPORTS_OSMESA)
+        set(FILAMENT_SUPPORTS_X11 FALSE)
+        add_definitions(-DFILAMENT_SUPPORTS_OSMESA)
     else ()
         if (FILAMENT_SUPPORTS_XCB)
             add_definitions(-DFILAMENT_SUPPORTS_XCB)

README.md

@@ -31,7 +31,7 @@ repositories {
 }
 
 dependencies {
-    implementation 'com.google.android.filament:filament-android:1.52.2'
+    implementation 'com.google.android.filament:filament-android:1.55.1'
 }
 ```
 
@@ -51,19 +51,9 @@ Here are all the libraries available in the group `com.google.android.filament`:
 iOS projects can use CocoaPods to install the latest release:
 
 ```shell
-pod 'Filament', '~> 1.52.2'
+pod 'Filament', '~> 1.55.1'
 ```
 
-### Snapshots
-
-If you prefer to live on the edge, you can download a continuous build by following the following
-steps:
-
-1. Find the [commit](https://github.com/google/filament/commits/main) you're interested in.
-2. Click the green check mark under the commit message.
-3. Click on the _Details_ link for the platform you're interested in.
-4. On the top left click _Summary_, then in the _Artifacts_ section choose the desired artifact.
-
 ## Documentation
 
 - [Filament](https://google.github.io/filament/Filament.html), an in-depth explanation of
@@ -176,6 +166,7 @@ steps:
   - [x] KHR_materials_unlit
   - [x] KHR_materials_variants
   - [x] KHR_materials_volume
+  - [x] KHR_materials_specular
   - [x] KHR_mesh_quantization
   - [x] KHR_texture_basisu
   - [x] KHR_texture_transform

RELEASE_NOTES.md

@@ -7,6 +7,56 @@ A new header is inserted each time a *tag* is created.
 Instead, if you are authoring a PR for the main branch, add your release note to
 [NEW_RELEASE_NOTES.md](./NEW_RELEASE_NOTES.md).
 
+## v1.55.1
+
+
+## v1.55.0
+- Add descriptor sets to describe shader resources. [⚠️ **New Material Version**]
+
+## v1.54.5
+
+
+## v1.54.4
+
+- Add support for multi-layered render target with array textures.
+
+## v1.54.3
+
+
+## v1.54.2
+
+- Add a `name` API to Filament objects for debugging handle use-after-free assertions
+
+## v1.54.1
+
+
+## v1.54.0
+
+- materials: add a new `stereoscopicType` material parameter. [⚠️ **New Material Version**]
+- Fix a crash when compiling shaders on IMG devices
+
+## v1.53.5
+
+- engine: Fix bug causing certain sampler parameters to not be applied correctly in GLES 2.0 and on
+  certain GLES 3.0 drivers.
+
+## v1.53.4
+
+
+## v1.53.3
+
+- Add drag and drop support for IBL files for desktop gltf_viewer.
+
+## v1.53.2
+
+
+## v1.53.1
+
+
+## v1.53.0
+
+- engine: fix skinning normals with large transforms (b/342459864) [⚠️ **New Material Version**]
+
 ## v1.52.3
 
 

android/build.gradle

@@ -83,12 +83,12 @@ buildscript {
         'minSdk': 21,
         'targetSdk': 34,
         'compileSdk': 34,
-        'kotlin': '1.9.21',
-        'kotlin_coroutines': '1.7.3',
+        'kotlin': '2.0.0',
+        'kotlin_coroutines': '1.9.0-RC',
         'buildTools': '34.0.0',
-        'ndk': '26.1.10909125',
-        'androidx_core': '1.12.0',
-        'androidx_annotations': '1.7.0'
+        'ndk': '27.0.11718014',
+        'androidx_core': '1.13.1',
+        'androidx_annotations': '1.8.0'
     ]
 
     ext.deps = [
@@ -104,7 +104,7 @@ buildscript {
     ]
 
     dependencies {
-        classpath 'com.android.tools.build:gradle:8.2.0'
+        classpath 'com.android.tools.build:gradle:8.4.1'
         classpath "org.jetbrains.kotlin:kotlin-gradle-plugin:${versions.kotlin}"
     }
 

android/filament-android/src/main/cpp/Engine.cpp

@@ -420,6 +420,13 @@ Java_com_google_android_filament_Engine_nSetPaused(JNIEnv*, jclass,
     engine->setPaused(paused);
 }
 
+extern "C" JNIEXPORT void JNICALL
+Java_com_google_android_filament_Engine_nUnprotected(JNIEnv*, jclass,
+        jlong nativeEngine, jboolean paused) {
+    Engine* engine = (Engine*) nativeEngine;
+    engine->unprotected();
+}
+
 // Managers...
 
 extern "C" JNIEXPORT jlong JNICALL
@@ -513,13 +520,12 @@ extern "C" JNIEXPORT void JNICALL Java_com_google_android_filament_Engine_nSetBu
 extern "C" JNIEXPORT void JNICALL Java_com_google_android_filament_Engine_nSetBuilderConfig(JNIEnv*,
         jclass, jlong nativeBuilder, jlong commandBufferSizeMB, jlong perRenderPassArenaSizeMB,
         jlong driverHandleArenaSizeMB, jlong minCommandBufferSizeMB, jlong perFrameCommandsSizeMB,
-        jlong jobSystemThreadCount,
-        jlong textureUseAfterFreePoolSize, jboolean disableParallelShaderCompile,
+        jlong jobSystemThreadCount, jboolean disableParallelShaderCompile,
         jint stereoscopicType, jlong stereoscopicEyeCount,
         jlong resourceAllocatorCacheSizeMB, jlong resourceAllocatorCacheMaxAge,
         jboolean disableHandleUseAfterFreeCheck,
         jint preferredShaderLanguage,
-        jboolean forceGLES2Context) {
+        jboolean forceGLES2Context, jboolean assertNativeWindowIsValid) {
     Engine::Builder* builder = (Engine::Builder*) nativeBuilder;
     Engine::Config config = {
             .commandBufferSizeMB = (uint32_t) commandBufferSizeMB,
@@ -528,7 +534,6 @@ extern "C" JNIEXPORT void JNICALL Java_com_google_android_filament_Engine_nSetBu
             .minCommandBufferSizeMB = (uint32_t) minCommandBufferSizeMB,
             .perFrameCommandsSizeMB = (uint32_t) perFrameCommandsSizeMB,
             .jobSystemThreadCount = (uint32_t) jobSystemThreadCount,
-            .textureUseAfterFreePoolSize = (uint32_t) textureUseAfterFreePoolSize,
             .disableParallelShaderCompile = (bool) disableParallelShaderCompile,
             .stereoscopicType = (Engine::StereoscopicType) stereoscopicType,
             .stereoscopicEyeCount = (uint8_t) stereoscopicEyeCount,
@@ -537,6 +542,7 @@ extern "C" JNIEXPORT void JNICALL Java_com_google_android_filament_Engine_nSetBu
             .disableHandleUseAfterFreeCheck = (bool) disableHandleUseAfterFreeCheck,
             .preferredShaderLanguage = (Engine::Config::ShaderLanguage) preferredShaderLanguage,
             .forceGLES2Context = (bool) forceGLES2Context,
+            .assertNativeWindowIsValid = (bool) assertNativeWindowIsValid,
     };
     builder->config(&config);
 }
@@ -564,3 +570,9 @@ Java_com_google_android_filament_Engine_nBuilderBuild(JNIEnv*, jclass, jlong nat
     Engine::Builder* builder = (Engine::Builder*) nativeBuilder;
     return (jlong) builder->build();
 }
+
+extern "C"
+JNIEXPORT jlong JNICALL
+Java_com_google_android_filament_Engine_getSteadyClockTimeNano(JNIEnv *env, jclass clazz) {
+    return (jlong)Engine::getSteadyClockTimeNano();
+}

android/filament-android/src/main/cpp/Material.cpp

@@ -25,12 +25,17 @@ using namespace filament;
 
 extern "C" JNIEXPORT jlong JNICALL
 Java_com_google_android_filament_Material_nBuilderBuild(JNIEnv *env, jclass,
-        jlong nativeEngine, jobject buffer_, jint size) {
+        jlong nativeEngine, jobject buffer_, jint size, jint shBandCount) {
     Engine* engine = (Engine*) nativeEngine;
     AutoBuffer buffer(env, buffer_, size);
-    Material* material = Material::Builder()
+    auto builder = Material::Builder();
+    if (shBandCount) {
+        builder.sphericalHarmonicsBandCount(shBandCount);
+    }
+    Material* material = builder
             .package(buffer.getData(), buffer.getSize())
             .build(*engine);
+
     return (jlong) material;
 }
 

android/filament-android/src/main/cpp/RenderableManager.cpp

@@ -253,16 +253,10 @@ Java_com_google_android_filament_RenderableManager_nBuilderMorphingStandard(JNIE
 }
 
 extern "C" JNIEXPORT void JNICALL
-Java_com_google_android_filament_RenderableManager_nBuilderSetMorphTargetBufferAt(JNIEnv*, jclass,
-        jlong nativeBuilder, int level, int primitiveIndex, jlong nativeMorphTargetBuffer,
-        int offset, int count) {
+Java_com_google_android_filament_RenderableManager_nBuilderSetMorphTargetBufferOffsetAt(JNIEnv*, jclass,
+        jlong nativeBuilder, int level, int primitiveIndex, int offset) {
     RenderableManager::Builder *builder = (RenderableManager::Builder *) nativeBuilder;
-    if (nativeMorphTargetBuffer) {
-        MorphTargetBuffer *morphTargetBuffer = (MorphTargetBuffer *) nativeMorphTargetBuffer;
-        builder->morphing(level, primitiveIndex, morphTargetBuffer, offset, count);
-    } else {
-        builder->morphing(level, primitiveIndex, offset, count);
-    }
+    builder->morphing(level, primitiveIndex, offset);
 }
 
 extern "C" JNIEXPORT void JNICALL
@@ -334,18 +328,12 @@ Java_com_google_android_filament_RenderableManager_nSetMorphWeights(JNIEnv* env,
 }
 
 extern "C" JNIEXPORT void JNICALL
-Java_com_google_android_filament_RenderableManager_nSetMorphTargetBufferAt(JNIEnv*,
+Java_com_google_android_filament_RenderableManager_nSetMorphTargetBufferOffsetAt(JNIEnv*,
         jclass, jlong nativeRenderableManager, jint i, int level, jint primitiveIndex,
-        jlong nativeMorphTargetBuffer, jint offset, jint count) {
+        jlong, jint offset) {
     RenderableManager *rm = (RenderableManager *) nativeRenderableManager;
-    if (nativeMorphTargetBuffer) {
-        MorphTargetBuffer *morphTargetBuffer = (MorphTargetBuffer *) nativeMorphTargetBuffer;
-        rm->setMorphTargetBufferAt((RenderableManager::Instance) i, (uint8_t) level,
-                (size_t) primitiveIndex, morphTargetBuffer, (size_t) offset, (size_t) count);
-    } else {
-        rm->setMorphTargetBufferAt((RenderableManager::Instance) i, (uint8_t) level,
-                (size_t) primitiveIndex, (size_t) offset, (size_t) count);
-    }
+    rm->setMorphTargetBufferOffsetAt((RenderableManager::Instance) i, (uint8_t) level,
+            (size_t) primitiveIndex, (size_t) offset);
 }
 
 extern "C" JNIEXPORT jint JNICALL

android/filament-android/src/main/cpp/Renderer.cpp

@@ -28,6 +28,14 @@
 using namespace filament;
 using namespace backend;
 
+
+extern "C" JNIEXPORT void JNICALL
+Java_com_google_android_filament_Renderer_nSkipFrame(JNIEnv *, jclass, jlong nativeRenderer,
+        jlong vsyncSteadyClockTimeNano) {
+    Renderer *renderer = (Renderer *) nativeRenderer;
+    renderer->skipFrame(uint64_t(vsyncSteadyClockTimeNano));
+}
+
 extern "C" JNIEXPORT jboolean JNICALL
 Java_com_google_android_filament_Renderer_nBeginFrame(JNIEnv *, jclass, jlong nativeRenderer,
         jlong nativeSwapChain, jlong frameTimeNanos) {
@@ -187,3 +195,10 @@ Java_com_google_android_filament_Renderer_nSetPresentationTime(JNIEnv *, jclass
     Renderer *renderer = (Renderer *) nativeRenderer;
     renderer->setPresentationTime(monotonicClockNanos);
 }
+
+extern "C" JNIEXPORT void JNICALL
+Java_com_google_android_filament_Renderer_nSetVsyncTime(JNIEnv *, jclass,
+    jlong nativeRenderer, jlong steadyClockTimeNano) {
+    Renderer *renderer = (Renderer *) nativeRenderer;
+    renderer->setVsyncTime(steadyClockTimeNano);
+}

android/filament-android/src/main/cpp/View.cpp

@@ -531,3 +531,12 @@ Java_com_google_android_filament_View_nGetFogEntity(JNIEnv *env, jclass clazz,
     View *view = (View *) nativeView;
     return (jint)view->getFogEntity().getId();
 }
+
+extern "C"
+JNIEXPORT void JNICALL
+Java_com_google_android_filament_View_nClearFrameHistory(JNIEnv *env, jclass clazz,
+        jlong nativeView, jlong nativeEngine) {
+    View *view = (View *) nativeView;
+    Engine *engine = (Engine *) nativeEngine;
+    view->clearFrameHistory(*engine);
+}

android/filament-android/src/main/java/com/google/android/filament/Engine.java

@@ -224,13 +224,12 @@ public class Engine {
             nSetBuilderConfig(mNativeBuilder, config.commandBufferSizeMB,
                     config.perRenderPassArenaSizeMB, config.driverHandleArenaSizeMB,
                     config.minCommandBufferSizeMB, config.perFrameCommandsSizeMB,
-                    config.jobSystemThreadCount,
-                    config.textureUseAfterFreePoolSize, config.disableParallelShaderCompile,
+                    config.jobSystemThreadCount, config.disableParallelShaderCompile,
                     config.stereoscopicType.ordinal(), config.stereoscopicEyeCount,
                     config.resourceAllocatorCacheSizeMB, config.resourceAllocatorCacheMaxAge,
                     config.disableHandleUseAfterFreeCheck,
                     config.preferredShaderLanguage.ordinal(),
-                    config.forceGLES2Context);
+                    config.forceGLES2Context, config.assertNativeWindowIsValid);
             return this;
         }
 
@@ -419,22 +418,25 @@ public class Engine {
          */
         public long stereoscopicEyeCount = 2;
 
-        /*
+        /**
          * @Deprecated This value is no longer used.
          */
         public long resourceAllocatorCacheSizeMB = 64;
 
-        /*
-         * This value determines for how many frames are texture entries kept in the cache.
+        /**
+         * This value determines how many frames texture entries are kept for in the cache. This
+         * is a soft limit, meaning some texture older than this are allowed to stay in the cache.
+         * Typically only one texture is evicted per frame.
+         * The default is 1.
          */
-        public long resourceAllocatorCacheMaxAge = 2;
+        public long resourceAllocatorCacheMaxAge = 1;
 
-        /*
+        /**
          * Disable backend handles use-after-free checks.
          */
         public boolean disableHandleUseAfterFreeCheck = false;
 
-        /*
+        /**
          * Sets a preferred shader language for Filament to use.
          *
          * The Metal backend supports two shader languages: MSL (Metal Shading Language) and
@@ -456,12 +458,19 @@ public class Engine {
         };
         public ShaderLanguage preferredShaderLanguage = ShaderLanguage.DEFAULT;
 
-        /*
+        /**
          * When the OpenGL ES backend is used, setting this value to true will force a GLES2.0
          * context if supported by the Platform, or if not, will have the backend pretend
          * it's a GLES2 context. Ignored on other backends.
          */
         public boolean forceGLES2Context = false;
+
+        /**
+         * Assert the native window associated to a SwapChain is valid when calling makeCurrent().
+         * This is only supported for:
+         *      - PlatformEGLAndroid
+         */
+        public boolean assertNativeWindowIsValid = false;
     }
 
     private Engine(long nativeEngine, Config config) {
@@ -1289,6 +1298,24 @@ public class Engine {
         nSetPaused(getNativeObject(), paused);
     }
 
+    /**
+     * Switch the command queue to unprotected mode. Protected mode can be activated via
+     * Renderer::beginFrame() using a protected SwapChain.
+     * @see Renderer
+     * @see SwapChain
+     */
+    public void unprotected() {
+        nUnprotected(getNativeObject());
+    }
+
+    /**
+     * Get the current time. This is a convenience function that simply returns the
+     * time in nanosecond since epoch of std::chrono::steady_clock.
+     * @return current time in nanosecond since epoch of std::chrono::steady_clock.
+     * @see Renderer#beginFrame
+     */
+    public static native long getSteadyClockTimeNano();
+
     @UsedByReflection("TextureHelper.java")
     public long getNativeObject() {
         if (mNativeObject == 0) {
@@ -1366,6 +1393,7 @@ public class Engine {
     private static native void nFlush(long nativeEngine);
     private static native boolean nIsPaused(long nativeEngine);
     private static native void nSetPaused(long nativeEngine, boolean paused);
+    private static native void nUnprotected(long nativeEngine);
     private static native long nGetTransformManager(long nativeEngine);
     private static native long nGetLightManager(long nativeEngine);
     private static native long nGetRenderableManager(long nativeEngine);
@@ -1384,12 +1412,11 @@ public class Engine {
     private static native void nSetBuilderConfig(long nativeBuilder, long commandBufferSizeMB,
             long perRenderPassArenaSizeMB, long driverHandleArenaSizeMB,
             long minCommandBufferSizeMB, long perFrameCommandsSizeMB, long jobSystemThreadCount,
-            long textureUseAfterFreePoolSize, boolean disableParallelShaderCompile,
-            int stereoscopicType, long stereoscopicEyeCount,
+            boolean disableParallelShaderCompile, int stereoscopicType, long stereoscopicEyeCount,
             long resourceAllocatorCacheSizeMB, long resourceAllocatorCacheMaxAge,
             boolean disableHandleUseAfterFreeCheck,
             int preferredShaderLanguage,
-            boolean forceGLES2Context);
+            boolean forceGLES2Context, boolean assertNativeWindowIsValid);
     private static native void nSetBuilderFeatureLevel(long nativeBuilder, int ordinal);
     private static native void nSetBuilderSharedContext(long nativeBuilder, long sharedContext);
     private static native void nSetBuilderPaused(long nativeBuilder, boolean paused);

android/filament-android/src/main/java/com/google/android/filament/Material.java

@@ -346,6 +346,7 @@ public class Material {
     public static class Builder {
         private Buffer mBuffer;
         private int mSize;
+        private int mShBandCount = 0;
 
         /**
          * Specifies the material data. The material data is a binary blob produced by
@@ -361,6 +362,22 @@ public class Material {
             return this;
         }
 
+        /**
+         * Sets the quality of the indirect lights computations. This is only taken into account
+         * if this material is lit and in the surface domain. This setting will affect the
+         * IndirectLight computation if one is specified on the Scene and Spherical Harmonics
+         * are used for the irradiance.
+         *
+         * @param shBandCount Number of spherical harmonic bands. Must be 1, 2 or 3 (default).
+         * @return Reference to this Builder for chaining calls.
+         * @see IndirectLight
+         */
+        @NonNull
+        public Builder sphericalHarmonicsBandCount(@IntRange(from = 0) int shBandCount) {
+            mShBandCount = shBandCount;
+            return this;
+        }
+
         /**
          * Creates and returns the Material object.
          *
@@ -372,7 +389,8 @@ public class Material {
          */
         @NonNull
         public Material build(@NonNull Engine engine) {
-            long nativeMaterial = nBuilderBuild(engine.getNativeObject(), mBuffer, mSize);
+            long nativeMaterial = nBuilderBuild(engine.getNativeObject(),
+                mBuffer, mSize, mShBandCount);
             if (nativeMaterial == 0) throw new IllegalStateException("Couldn't create Material");
             return new Material(nativeMaterial);
         }
@@ -1023,7 +1041,7 @@ public class Material {
         mNativeObject = 0;
     }
 
-    private static native long nBuilderBuild(long nativeEngine, @NonNull Buffer buffer, int size);
+    private static native long nBuilderBuild(long nativeEngine, @NonNull Buffer buffer, int size, int shBandCount);
     private static native long nCreateInstance(long nativeMaterial);
     private static native long nCreateInstanceWithName(long nativeMaterial, @NonNull String name);
     private static native long nGetDefaultInstance(long nativeMaterial);

android/filament-android/src/main/java/com/google/android/filament/MorphTargetBuffer.java

@@ -74,7 +74,7 @@ public class MorphTargetBuffer {
          *
          * @exception IllegalStateException if the MorphTargetBuffer could not be created
          *
-         * @see #setMorphTargetBufferAt
+         * @see #setMorphTargetBufferOffsetAt
          */
         @NonNull
         public MorphTargetBuffer build(@NonNull Engine engine) {

android/filament-android/src/main/java/com/google/android/filament/RenderableManager.java

@@ -569,40 +569,13 @@ public class RenderableManager {
          *
          * @param level the level of detail (lod), only 0 can be specified
          * @param primitiveIndex zero-based index of the primitive, must be less than the count passed to Builder constructor
-         * @param morphTargetBuffer specifies the morph target buffer
          * @param offset specifies where in the morph target buffer to start reading (expressed as a number of vertices)
-         * @param count number of vertices in the morph target buffer to read, must equal the geometry's count (for triangles, this should be a multiple of 3)
          */
         @NonNull
         public Builder morphing(@IntRange(from = 0) int level,
                                 @IntRange(from = 0) int primitiveIndex,
-                                @IntRange(from = 0) int offset,
-                                @IntRange(from = 0) int count) {
-            nBuilderSetMorphTargetBufferAt(mNativeBuilder, level, primitiveIndex, 0, offset, count);
-            return this;
-        }
-
-        /** @deprecated */
-        @Deprecated
-        @NonNull
-        public Builder morphing(@IntRange(from = 0) int level,
-                                @IntRange(from = 0) int primitiveIndex,
-                                @NonNull MorphTargetBuffer morphTargetBuffer,
-                                @IntRange(from = 0) int offset,
-                                @IntRange(from = 0) int count) {
-            nBuilderSetMorphTargetBufferAt(mNativeBuilder, level, primitiveIndex,
-                    morphTargetBuffer.getNativeObject(), offset, count);
-            return this;
-        }
-
-        /** @deprecated */
-        @Deprecated
-        @NonNull
-        public Builder morphing(@IntRange(from = 0) int level,
-                                @IntRange(from = 0) int primitiveIndex,
-                                @NonNull MorphTargetBuffer morphTargetBuffer) {
-            nBuilderSetMorphTargetBufferAt(mNativeBuilder, level, primitiveIndex,
-                    morphTargetBuffer.getNativeObject(), 0, morphTargetBuffer.getVertexCount());
+                                @IntRange(from = 0) int offset) {
+            nBuilderSetMorphTargetBufferOffsetAt(mNativeBuilder, level, primitiveIndex, offset);
             return this;
         }
 
@@ -705,34 +678,11 @@ public class RenderableManager {
      *
      * @see Builder#morphing
      */
-    public void setMorphTargetBufferAt(@EntityInstance int i,
+    public void setMorphTargetBufferOffsetAt(@EntityInstance int i,
                                        @IntRange(from = 0) int level,
                                        @IntRange(from = 0) int primitiveIndex,
-                                       @IntRange(from = 0) int offset,
-                                       @IntRange(from = 0) int count) {
-        nSetMorphTargetBufferAt(mNativeObject, i, level, primitiveIndex, 0, offset, count);
-    }
-
-    /** @deprecated */
-    @Deprecated
-    public void setMorphTargetBufferAt(@EntityInstance int i,
-                                       @IntRange(from = 0) int level,
-                                       @IntRange(from = 0) int primitiveIndex,
-                                       @NonNull MorphTargetBuffer morphTargetBuffer,
-                                       @IntRange(from = 0) int offset,
-                                       @IntRange(from = 0) int count) {
-        nSetMorphTargetBufferAt(mNativeObject, i, level, primitiveIndex,
-                morphTargetBuffer.getNativeObject(), offset, count);
-    }
-
-    /** @deprecated */
-    @Deprecated
-    public void setMorphTargetBufferAt(@EntityInstance int i,
-                                       @IntRange(from = 0) int level,
-                                       @IntRange(from = 0) int primitiveIndex,
-                                       @NonNull MorphTargetBuffer morphTargetBuffer) {
-        nSetMorphTargetBufferAt(mNativeObject, i, level, primitiveIndex,
-                morphTargetBuffer.getNativeObject(), 0, morphTargetBuffer.getVertexCount());
+                                       @IntRange(from = 0) int offset) {
+        nSetMorphTargetBufferOffsetAt(mNativeObject, i, level, primitiveIndex, 0, offset);
     }
 
     /**
@@ -1033,7 +983,7 @@ public class RenderableManager {
     private static native void nBuilderSkinningBuffer(long nativeBuilder, long nativeSkinningBuffer, int boneCount, int offset);
     private static native void nBuilderMorphing(long nativeBuilder, int targetCount);
     private static native void nBuilderMorphingStandard(long nativeBuilder, long nativeMorphTargetBuffer);
-    private static native void nBuilderSetMorphTargetBufferAt(long nativeBuilder, int level, int primitiveIndex, long nativeMorphTargetBuffer, int offset, int count);
+    private static native void nBuilderSetMorphTargetBufferOffsetAt(long nativeBuilder, int level, int primitiveIndex, int offset);
     private static native void nBuilderEnableSkinningBuffers(long nativeBuilder, boolean enabled);
     private static native void nBuilderFog(long nativeBuilder, boolean enabled);
     private static native void nBuilderLightChannel(long nativeRenderableManager, int channel, boolean enable);
@@ -1043,7 +993,7 @@ public class RenderableManager {
     private static native int nSetBonesAsMatrices(long nativeObject, int i, Buffer matrices, int remaining, int boneCount, int offset);
     private static native int nSetBonesAsQuaternions(long nativeObject, int i, Buffer quaternions, int remaining, int boneCount, int offset);
     private static native void nSetMorphWeights(long nativeObject, int instance, float[] weights, int offset);
-    private static native void nSetMorphTargetBufferAt(long nativeObject, int i, int level, int primitiveIndex, long nativeMorphTargetBuffer, int offset, int count);
+    private static native void nSetMorphTargetBufferOffsetAt(long nativeObject, int i, int level, int primitiveIndex, long nativeMorphTargetBuffer, int offset);
     private static native int nGetMorphTargetCount(long nativeObject, int i);
     private static native void nSetAxisAlignedBoundingBox(long nativeRenderableManager, int i, float cx, float cy, float cz, float ex, float ey, float ez);
     private static native void nSetLayerMask(long nativeRenderableManager, int i, int select, int value);

android/filament-android/src/main/java/com/google/android/filament/Renderer.java

@@ -284,6 +284,33 @@ public class Renderer {
         nSetPresentationTime(getNativeObject(), monotonicClockNanos);
     }
 
+    /**
+     * The use of this method is optional. It sets the VSYNC time expressed as the duration in
+     * nanosecond since epoch of std::chrono::steady_clock.
+     * If called, passing 0 to frameTimeNanos in Renderer.BeginFrame will use this
+     * time instead.
+     * @param steadyClockTimeNano duration in nanosecond since epoch of std::chrono::steady_clock
+     * @see Engine#getSteadyClockTimeNano
+     * @see Renderer#beginFrame
+     */
+    public void setVsyncTime(long steadyClockTimeNano) {
+        nSetVsyncTime(getNativeObject(), steadyClockTimeNano);
+    }
+
+    /**
+     * Call skipFrame when momentarily skipping frames, for instance if the content of the
+     * scene doesn't change.
+     *
+     * @param vsyncSteadyClockTimeNano The time in nanoseconds when the frame started being rendered,
+     *                       in the {@link System#nanoTime()} timebase. Divide this value by 1000000 to
+     *                       convert it to the {@link android.os.SystemClock#uptimeMillis()}
+     *                       time base. This typically comes from
+     *                       {@link android.view.Choreographer.FrameCallback}.
+     */
+    public void skipFrame(long vsyncSteadyClockTimeNano) {
+        nSkipFrame(getNativeObject(), vsyncSteadyClockTimeNano);
+    }
+
     /**
      * Sets up a frame for this <code>Renderer</code>.
      * <p><code>beginFrame</code> manages frame pacing, and returns whether or not a frame should be
@@ -702,6 +729,8 @@ public class Renderer {
     }
 
     private static native void nSetPresentationTime(long nativeObject, long monotonicClockNanos);
+    private static native void nSetVsyncTime(long nativeObject, long steadyClockTimeNano);
+    private static native void nSkipFrame(long nativeObject, long vsyncSteadyClockTimeNano);
     private static native boolean nBeginFrame(long nativeRenderer, long nativeSwapChain, long frameTimeNanos);
     private static native void nEndFrame(long nativeRenderer);
     private static native void nRender(long nativeRenderer, long nativeView);

android/filament-android/src/main/java/com/google/android/filament/View.java

@@ -1233,6 +1233,18 @@ public class View {
         return nGetFogEntity(getNativeObject());
     }
 
+    /**
+     * When certain temporal features are used (e.g.: TAA or Screen-space reflections), the view
+     * keeps a history of previous frame renders associated with the Renderer the view was last
+     * used with. When switching Renderer, it may be necessary to clear that history by calling
+     * this method. Similarly, if the whole content of the screen change, like when a cut-scene
+     * starts, clearing the history might be needed to avoid artifacts due to the previous frame
+     * being very different.
+     */
+    public void clearFrameHistory(Engine engine) {
+        nClearFrameHistory(getNativeObject(), engine.getNativeObject());
+    }
+
     public long getNativeObject() {
         if (mNativeObject == 0) {
             throw new IllegalStateException("Calling method on destroyed View");
@@ -1294,7 +1306,7 @@ public class View {
     private static native void nSetMaterialGlobal(long nativeView, int index, float x, float y, float z, float w);
     private static native void nGetMaterialGlobal(long nativeView, int index, float[] out);
     private static native int nGetFogEntity(long nativeView);
-
+    private static native void nClearFrameHistory(long nativeView, long nativeEngine);
 
     /**
      * List of available ambient occlusion techniques.

android/filament-utils-android/src/main/cpp/Manipulator.cpp

@@ -125,6 +125,11 @@ extern "C" JNIEXPORT void Java_com_google_android_filament_utils_Manipulator_nBu
     builder->groundPlane(a, b, c, d);
 }
 
+extern "C" JNIEXPORT void Java_com_google_android_filament_utils_Manipulator_nBuilderPanning(JNIEnv*, jclass, jlong nativeBuilder, jboolean enabled) {
+    Builder* builder = (Builder*) nativeBuilder;
+    builder->panning(enabled);
+}
+
 extern "C" JNIEXPORT long Java_com_google_android_filament_utils_Manipulator_nBuilderBuild(JNIEnv*, jclass, jlong nativeBuilder, jint mode) {
     Builder* builder = (Builder*) nativeBuilder;
     return (jlong) builder->build((Mode) mode);

android/filament-utils-android/src/main/java/com/google/android/filament/utils/Manipulator.java

@@ -274,6 +274,17 @@ public class Manipulator {
             return this;
         }
 
+        /**
+         * Sets whether panning is enabled in the manipulator.
+         *
+         * @return this <code>Builder</code> object for chaining calls
+         */
+        @NonNull
+        public Builder panning(Boolean enabled) {
+            nBuilderPanning(mNativeBuilder, enabled);
+            return this;
+        }
+
         /**
          * Creates and returns the <code>Manipulator</code> object.
          *
@@ -483,6 +494,7 @@ public class Manipulator {
     private static native void nBuilderFlightPanSpeed(long nativeBuilder, float x, float y);
     private static native void nBuilderFlightMoveDamping(long nativeBuilder, float damping);
     private static native void nBuilderGroundPlane(long nativeBuilder, float a, float b, float c, float d);
+    private static native void nBuilderPanning(long nativeBuilder, Boolean enabled);
     private static native long nBuilderBuild(long nativeBuilder, int mode);
 
     private static native void nDestroyManipulator(long nativeManip);

android/gradle.properties

@@ -1,5 +1,5 @@
 GROUP=com.google.android.filament
-VERSION_NAME=1.52.2
+VERSION_NAME=1.55.1
 
 POM_DESCRIPTION=Real-time physically based rendering engine for Android.
 

android/gradle/wrapper/gradle-wrapper.properties

@@ -1,6 +1,6 @@
 #Wed Nov 17 10:40:18 PST 2021
 distributionBase=GRADLE_USER_HOME
 distributionPath=wrapper/dists
-distributionUrl=https\://services.gradle.org/distributions/gradle-8.2-bin.zip
+distributionUrl=https\://services.gradle.org/distributions/gradle-8.6-bin.zip
 zipStoreBase=GRADLE_USER_HOME
 zipStorePath=wrapper/dists

build.sh

@@ -64,6 +64,9 @@ function print_help {
     echo "        enabling debug paths in the backend from the build script. For example, make a"
     echo "        systrace-enabled build without directly changing #defines. Remember to add -f when"
     echo "        changing this option."
+    echo "    -X osmesa_path"
+    echo "        Indicates a path to a completed OSMesa build. OSMesa is used to create an offscreen GL"
+    echo "        context for software rasterization"
     echo "    -S type"
     echo "        Enable stereoscopic rendering where type is one of [instanced|multiview]. This is only"
     echo "        meant for building the samples."
@@ -180,6 +183,8 @@ BACKEND_DEBUG_FLAG_OPTION=""
 
 STEREOSCOPIC_OPTION=""
 
+OSMESA_OPTION=""
+
 IOS_BUILD_SIMULATOR=false
 BUILD_UNIVERSAL_LIBRARIES=false
 
@@ -240,6 +245,7 @@ function build_desktop_target {
             ${ASAN_UBSAN_OPTION} \
             ${BACKEND_DEBUG_FLAG_OPTION} \
             ${STEREOSCOPIC_OPTION} \
+            ${OSMESA_OPTION} \
             ${architectures} \
             ../..
         ln -sf "out/cmake-${lc_target}/compile_commands.json" \
@@ -796,7 +802,7 @@ function check_debug_release_build {
 
 pushd "$(dirname "$0")" > /dev/null
 
-while getopts ":hacCfgijmp:q:uvslwedk:bx:S:" opt; do
+while getopts ":hacCfgijmp:q:uvslwedk:bx:S:X:" opt; do
     case ${opt} in
         h)
             print_help
@@ -950,6 +956,8 @@ while getopts ":hacCfgijmp:q:uvslwedk:bx:S:" opt; do
                     exit 1
             esac
             ;;
+        X)  OSMESA_OPTION="-DFILAMENT_OSMESA_PATH=${OPTARG}"
+            ;;
         \?)
             echo "Invalid option: -${OPTARG}" >&2
             echo ""

build/android/build.sh

@@ -57,15 +57,10 @@ FILAMENT_NDK_VERSION=${FILAMENT_NDK_VERSION:-$(cat `dirname $0`/ndk.version)}
 
 # Install the required NDK version specifically (if not present)
 if [[ ! -d "${ANDROID_HOME}/ndk/$FILAMENT_NDK_VERSION" ]]; then
-    ${ANDROID_HOME}/cmdline-tools/latest/bin/sdkmanager "ndk;$FILAMENT_NDK_VERSION" > /dev/null
+    yes | ${ANDROID_HOME}/cmdline-tools/latest/bin/sdkmanager --licenses
+    ${ANDROID_HOME}/cmdline-tools/latest/bin/sdkmanager "ndk;$FILAMENT_NDK_VERSION"
 fi
-
-# Only build 1 64 bit target during presubmit to cut down build times during presubmit
-# Continuous builds will build everything
 ANDROID_ABIS=
-if [[ "$TARGET" == "presubmit" ]]; then
-  ANDROID_ABIS="-q arm64-v8a"
-fi
 
 # Build the Android sample-gltf-viewer APK during release.
 BUILD_SAMPLES=
@@ -73,5 +68,19 @@ if [[ "$TARGET" == "release" ]]; then
     BUILD_SAMPLES="-k sample-gltf-viewer"
 fi
 
+function build_android() {
+    local ABI=$1
+
+    # Do the following in two steps so that we do not run out of space
+    if [[ -n "${BUILD_DEBUG}" ]]; then
+        FILAMENT_NDK_VERSION=${FILAMENT_NDK_VERSION} ./build.sh -p android -q ${ABI} -c ${BUILD_SAMPLES} ${GENERATE_ARCHIVES} ${BUILD_DEBUG}
+        rm -rf out/cmake-android-debug-*
+    fi
+    if [[ -n "${BUILD_RELEASE}" ]]; then
+        FILAMENT_NDK_VERSION=${FILAMENT_NDK_VERSION} ./build.sh -p android -q ${ABI} -c ${BUILD_SAMPLES} ${GENERATE_ARCHIVES} ${BUILD_RELEASE}
+        rm -rf out/cmake-android-release-*
+    fi
+}
+
 pushd `dirname $0`/../.. > /dev/null
-FILAMENT_NDK_VERSION=${FILAMENT_NDK_VERSION} ./build.sh -p android $ANDROID_ABIS -c $BUILD_SAMPLES $GENERATE_ARCHIVES $BUILD_DEBUG $BUILD_RELEASE
+build_android $2

build/android/ndk.version

@@ -1 +1 @@
-26.1.10909125
+27.0.11718014

build/web/ci-common.sh

@@ -1,7 +1,6 @@
 #!/bin/bash
 if [ `uname` == "Linux" ];then
-    curl -OL https://github.com/ninja-build/ninja/releases/download/v1.10.2/ninja-linux.zip
-    unzip -q ninja-linux.zip
+    source `dirname $0`/../linux/ci-common.sh
 elif [ `uname` == "Darwin" ];then
     curl -OL https://github.com/ninja-build/ninja/releases/download/v1.10.2/ninja-mac.zip
     unzip -q ninja-mac.zip
@@ -13,9 +12,6 @@ fi
 chmod +x ninja
 export PATH="$PWD:$PATH"
 
-# FIXME: kokoro machines have node and npm but currently they are symlinked to non-existent files
-# npm install -g typescript
-
 # Install emscripten.
 curl -L https://github.com/emscripten-core/emsdk/archive/refs/tags/3.1.15.zip > emsdk.zip
 unzip emsdk.zip ; mv emsdk-* emsdk ; cd emsdk

build/windows/build-github.bat

@@ -115,17 +115,23 @@ cmake ..\.. ^
     -DFILAMENT_SUPPORTS_VULKAN=ON ^
     || exit /b
 
+set build_flags=-j %NUMBER_OF_PROCESSORS%
+
+@echo on
+
+:: we've upgraded the windows machines, so the following are no longer accurate as of 09/19/24, but
+:: keeping around the comment for record.
+
 :: Attempt to fix "error C1060: compiler is out of heap space" seen on CI.
 :: Some resource libraries require significant heap space to compile, so first compile them serially.
-@echo on
-cmake --build . --target filagui --config %config% || exit /b
-cmake --build . --target uberarchive --config %config% || exit /b
-cmake --build . --target gltf-demo-resources --config %config% || exit /b
-cmake --build . --target filamentapp-resources --config %config% || exit /b
-cmake --build . --target sample-resources --config %config% || exit /b
-cmake --build . --target suzanne-resources --config %config% || exit /b
+:: cmake --build . --target filagui --config %config% %build_flags% || exit /b
+:: cmake --build . --target uberarchive --config %config% %build_flags% || exit /b
+:: cmake --build . --target gltf-demo-resources --config %config% %build_flags% || exit /b
+:: cmake --build . --target filamentapp-resources --config %config% %build_flags% || exit /b
+:: cmake --build . --target sample-resources --config %config% %build_flags% || exit /b
+:: cmake --build . --target suzanne-resources --config %config% %build_flags% || exit /b
 
-cmake --build . %INSTALL% --config %config% -- /m || exit /b
+cmake --build . %INSTALL% --config %config% %build_flags% -- /m || exit /b
 @echo off
 
 echo Disk info after building variant: %variant%

docs/Materials.md.html

@@ -1308,7 +1308,12 @@ Description
       declare a variable called `eyeDirection` you can access it in the fragment shader using
       `variable_eyeDirection`. In the vertex shader, the interpolant name is simply a member of
       the `MaterialVertexInputs` structure (`material.eyeDirection` in your example). Each
-      interpolant is of type `float4` (`vec4`) in the shaders.
+      interpolant is of type `float4` (`vec4`) in the shaders. By default the precision of the
+      interpolant is `highp` in *both* the vertex and fragment shaders.
+      An alternate syntax can be used to specify both the name and precision of the interpolant.
+      In this case the specified precision is used as-is in both fragment and vertex stages, in
+      particular if `default` is specified the default precision is used is the fragment shader
+      (`mediump`) and in the vertex shader (`highp`).
 
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ JSON
 material {
@@ -1320,7 +1325,11 @@ material {
         }
     ],
     variables : [
-         eyeDirection
+         eyeDirection,
+         {
+            name : eyeColor,
+            precision : medium
+         }
     ],
     vertexDomain : device,
     depthWrite : false,

filament/CMakeLists.txt

@@ -61,10 +61,12 @@ set(SRCS
         src/Engine.cpp
         src/Exposure.cpp
         src/Fence.cpp
+        src/FilamentBuilder.cpp
         src/FrameInfo.cpp
         src/FrameSkipper.cpp
         src/Froxelizer.cpp
         src/Frustum.cpp
+        src/HwDescriptorSetLayoutFactory.cpp
         src/HwRenderPrimitiveFactory.cpp
         src/HwVertexBufferInfoFactory.cpp
         src/IndexBuffer.cpp
@@ -75,8 +77,6 @@ set(SRCS
         src/MaterialInstance.cpp
         src/MaterialParser.cpp
         src/MorphTargetBuffer.cpp
-        src/PerViewUniforms.cpp
-        src/PerShadowMapUniforms.cpp
         src/PostProcessManager.cpp
         src/RenderPass.cpp
         src/RenderPrimitive.cpp
@@ -125,6 +125,12 @@ set(SRCS
         src/details/Texture.cpp
         src/details/VertexBuffer.cpp
         src/details/View.cpp
+        src/ds/ColorPassDescriptorSet.cpp
+        src/ds/DescriptorSet.cpp
+        src/ds/DescriptorSetLayout.cpp
+        src/ds/PostProcessDescriptorSet.cpp
+        src/ds/ShadowMapDescriptorSet.cpp
+        src/ds/SsrPassDescriptorSet.cpp
         src/fg/Blackboard.cpp
         src/fg/DependencyGraph.cpp
         src/fg/FrameGraph.cpp
@@ -148,23 +154,21 @@ set(PRIVATE_HDRS
         src/FrameInfo.h
         src/FrameSkipper.h
         src/Froxelizer.h
+        src/HwDescriptorSetLayoutFactory.h
         src/HwRenderPrimitiveFactory.h
         src/HwVertexBufferInfoFactory.h
         src/Intersections.h
         src/MaterialParser.h
-        src/PerViewUniforms.h
-        src/PerShadowMapUniforms.h
         src/PIDController.h
         src/PostProcessManager.h
-        src/RendererUtils.h
         src/RenderPass.h
         src/RenderPrimitive.h
+        src/RendererUtils.h
         src/ResourceAllocator.h
         src/ResourceList.h
         src/ShadowMap.h
         src/ShadowMapManager.h
         src/SharedHandle.h
-        src/TypedUniformBuffer.h
         src/UniformBuffer.h
         src/components/CameraManager.h
         src/components/LightManager.h
@@ -192,6 +196,14 @@ set(PRIVATE_HDRS
         src/details/Texture.h
         src/details/VertexBuffer.h
         src/details/View.h
+        src/downcast.h
+        src/ds/ColorPassDescriptorSet.h
+        src/ds/DescriptorSetLayout.h
+        src/ds/PostProcessDescriptorSet.h
+        src/ds/ShadowMapDescriptorSet.h
+        src/ds/SsrPassDescriptorSet.h
+        src/ds/TypedBuffer.h
+        src/ds/TypedUniformBuffer.h
         src/fg/Blackboard.h
         src/fg/FrameGraph.h
         src/fg/FrameGraphId.h
@@ -209,7 +221,6 @@ set(PRIVATE_HDRS
         src/materials/fsr/ffx_a.h
         src/materials/fsr/ffx_fsr1.h
         src/materials/fsr/ffx_fsr1_mobile.fs
-        src/downcast.h
 )
 
 set(MATERIAL_SRCS

filament/backend/CMakeLists.txt

@@ -12,6 +12,7 @@ set(PUBLIC_HDRS
         include/backend/AcquiredImage.h
         include/backend/BufferDescriptor.h
         include/backend/CallbackHandler.h
+        include/backend/DescriptorSetOffsetArray.h
         include/backend/DriverApiForward.h
         include/backend/DriverEnums.h
         include/backend/Handle.h
@@ -69,9 +70,13 @@ set(PRIVATE_HDRS
 if (FILAMENT_SUPPORTS_OPENGL AND NOT FILAMENT_USE_EXTERNAL_GLES3)
     list(APPEND SRCS
             include/backend/platforms/OpenGLPlatform.h
+            src/opengl/BindingMap.h
             src/opengl/gl_headers.cpp
             src/opengl/gl_headers.h
             src/opengl/GLBufferObject.h
+            src/opengl/GLDescriptorSet.cpp
+            src/opengl/GLDescriptorSet.h
+            src/opengl/GLDescriptorSetLayout.h
             src/opengl/GLTexture.h
             src/opengl/GLUtils.cpp
             src/opengl/GLUtils.h
@@ -114,6 +119,8 @@ if (FILAMENT_SUPPORTS_OPENGL AND NOT FILAMENT_USE_EXTERNAL_GLES3)
             list(APPEND SRCS src/opengl/platforms/PlatformGLX.cpp)
         elseif (FILAMENT_SUPPORTS_EGL_ON_LINUX)
             list(APPEND SRCS src/opengl/platforms/PlatformEGLHeadless.cpp)
+        elseif (FILAMENT_SUPPORTS_OSMESA)
+            list(APPEND SRCS src/opengl/platforms/PlatformOSMesa.cpp)
         endif()
     elseif (WIN32)
         list(APPEND SRCS src/opengl/platforms/PlatformWGL.cpp)
@@ -361,6 +368,11 @@ set(LINUX_LINKER_OPTIMIZATION_FLAGS
         -Wl,--exclude-libs,bluegl
 )
 
+if (LINUX AND FILAMENT_SUPPORTS_OSMESA)
+    set(OSMESA_COMPILE_FLAGS
+        -I${FILAMENT_OSMESA_PATH}/include/GL)
+endif()
+
 if (MSVC)
     set(FILAMENT_WARNINGS /W3)
 else()
@@ -381,6 +393,7 @@ endif()
 
 target_compile_options(${TARGET} PRIVATE
         ${FILAMENT_WARNINGS}
+        ${OSMESA_COMPILE_FLAGS}
         $<$<CONFIG:Release>:${OPTIMIZATION_FLAGS}>
         $<$<AND:$<PLATFORM_ID:Darwin>,$<CONFIG:Release>>:${DARWIN_OPTIMIZATION_FLAGS}>
 )
@@ -390,6 +403,7 @@ if (FILAMENT_SUPPORTS_METAL)
 endif()
 
 target_link_libraries(${TARGET} PRIVATE
+        ${OSMESA_LINKER_FLAGS}
         $<$<AND:$<PLATFORM_ID:Linux>,$<CONFIG:Release>>:${LINUX_LINKER_OPTIMIZATION_FLAGS}>
 )
 
@@ -424,6 +438,7 @@ if (APPLE OR LINUX)
         test/test_StencilBuffer.cpp
         test/test_Scissor.cpp
         test/test_MipLevels.cpp
+        test/test_Handles.cpp
     )
     set(BACKEND_TEST_LIBS
         backend
@@ -495,21 +510,40 @@ endif()
 
 # ==================================================================================================
 # Compute tests
+#
+#if (NOT IOS AND NOT WEBGL)
+#
+#add_executable(compute_test
+#        test/ComputeTest.cpp
+#        test/Arguments.cpp
+#        test/test_ComputeBasic.cpp
+#        )
+#
+#target_link_libraries(compute_test PRIVATE
+#        backend
+#        getopt
+#        gtest
+#        )
+#
+#set_target_properties(compute_test PROPERTIES FOLDER Tests)
+#
+#endif()
 
-if (NOT IOS AND NOT WEBGL)
+# ==================================================================================================
+# Metal utils tests
 
-add_executable(compute_test
-        test/ComputeTest.cpp
-        test/Arguments.cpp
-        test/test_ComputeBasic.cpp
-        )
+if (APPLE AND NOT IOS)
 
-target_link_libraries(compute_test PRIVATE
+add_executable(metal_utils_test test/MetalTest.mm)
+
+target_compile_options(metal_utils_test PRIVATE "-fobjc-arc")
+
+target_link_libraries(metal_utils_test PRIVATE
         backend
         getopt
         gtest
         )
 
-set_target_properties(compute_test PROPERTIES FOLDER Tests)
+set_target_properties(metal_utils_test PROPERTIES FOLDER Tests)
 
 endif()

filament/backend/include/backend/DescriptorSetOffsetArray.h

@@ -0,0 +1,101 @@
+/*
+ * Copyright (C) 2024 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef TNT_FILAMENT_BACKEND_COMMANDSTREAMVECTOR_H
+#define TNT_FILAMENT_BACKEND_COMMANDSTREAMVECTOR_H
+
+#include <backend/DriverApiForward.h>
+
+#include <initializer_list>
+#include <memory>
+
+#include <stddef.h>
+#include <stdint.h>
+
+
+namespace filament::backend {
+
+void* allocateFromCommandStream(DriverApi& driver, size_t size, size_t alignment) noexcept;
+
+class DescriptorSetOffsetArray {
+public:
+    using value_type = uint32_t;
+    using reference = value_type&;
+    using const_reference = value_type const&;
+    using size_type = uint32_t;
+    using difference_type = int32_t;
+    using pointer = value_type*;
+    using const_pointer = value_type const*;
+    using iterator = pointer;
+    using const_iterator = const_pointer;
+
+    DescriptorSetOffsetArray() noexcept = default;
+
+    ~DescriptorSetOffsetArray() noexcept = default;
+
+    DescriptorSetOffsetArray(size_type size, DriverApi& driver) noexcept {
+        mOffsets = (value_type *)allocateFromCommandStream(driver,
+                size * sizeof(value_type), alignof(value_type));
+        std::uninitialized_fill_n(mOffsets, size, 0);
+    }
+
+    DescriptorSetOffsetArray(std::initializer_list<uint32_t> list, DriverApi& driver) noexcept {
+        mOffsets = (value_type *)allocateFromCommandStream(driver,
+                list.size() * sizeof(value_type), alignof(value_type));
+        std::uninitialized_copy(list.begin(), list.end(), mOffsets);
+    }
+
+    DescriptorSetOffsetArray(DescriptorSetOffsetArray const&) = delete;
+    DescriptorSetOffsetArray& operator=(DescriptorSetOffsetArray const&) = delete;
+
+    DescriptorSetOffsetArray(DescriptorSetOffsetArray&& rhs) noexcept
+            : mOffsets(rhs.mOffsets) {
+        rhs.mOffsets = nullptr;
+    }
+
+    DescriptorSetOffsetArray& operator=(DescriptorSetOffsetArray&& rhs) noexcept {
+        if (this != &rhs) {
+            mOffsets = rhs.mOffsets;
+            rhs.mOffsets = nullptr;
+        }
+        return *this;
+    }
+
+    bool empty() const noexcept { return mOffsets == nullptr; }
+
+    value_type* data() noexcept { return mOffsets; }
+    const value_type* data() const noexcept { return mOffsets; }
+
+
+    reference operator[](size_type n) noexcept {
+        return *(data() + n);
+    }
+
+    const_reference operator[](size_type n) const noexcept {
+        return *(data() + n);
+    }
+
+    void clear() noexcept {
+        mOffsets = nullptr;
+    }
+
+private:
+    value_type *mOffsets = nullptr;
+};
+
+} // namespace filament::backend
+
+#endif //TNT_FILAMENT_BACKEND_COMMANDSTREAMVECTOR_H

filament/backend/include/backend/DriverEnums.h

@@ -19,13 +19,16 @@
 #ifndef TNT_FILAMENT_BACKEND_DRIVERENUMS_H
 #define TNT_FILAMENT_BACKEND_DRIVERENUMS_H
 
-#include <utils/BitmaskEnum.h>
 #include <utils/unwindows.h> // Because we define ERROR in the FenceStatus enum.
 
 #include <backend/Platform.h>
 #include <backend/PresentCallable.h>
 
+#include <utils/BitmaskEnum.h>
+#include <utils/FixedCapacityVector.h>
 #include <utils/Invocable.h>
+#include <utils/compiler.h>
+#include <utils/debug.h>
 #include <utils/ostream.h>
 
 #include <math/vec4.h>
@@ -97,6 +100,8 @@ static constexpr size_t MAX_VERTEX_ATTRIBUTE_COUNT  = 16;   // This is guarantee
 static constexpr size_t MAX_SAMPLER_COUNT           = 62;   // Maximum needed at feature level 3.
 static constexpr size_t MAX_VERTEX_BUFFER_COUNT     = 16;   // Max number of bound buffer objects.
 static constexpr size_t MAX_SSBO_COUNT              = 4;    // This is guaranteed by OpenGL ES.
+static constexpr size_t MAX_DESCRIPTOR_SET_COUNT    = 4;    // This is guaranteed by Vulkan.
+static constexpr size_t MAX_DESCRIPTOR_COUNT        = 64;   // per set
 
 static constexpr size_t MAX_PUSH_CONSTANT_COUNT     = 32;   // Vulkan 1.1 spec allows for 128-byte
                                                             // of push constant (we assume 4-byte
@@ -191,6 +196,70 @@ static constexpr const char* shaderLanguageToString(ShaderLanguage shaderLanguag
     }
 }
 
+enum class ShaderStage : uint8_t {
+    VERTEX = 0,
+    FRAGMENT = 1,
+    COMPUTE = 2
+};
+
+static constexpr size_t PIPELINE_STAGE_COUNT = 2;
+enum class ShaderStageFlags : uint8_t {
+    NONE        =    0,
+    VERTEX      =    0x1,
+    FRAGMENT    =    0x2,
+    COMPUTE     =    0x4,
+    ALL_SHADER_STAGE_FLAGS = VERTEX | FRAGMENT | COMPUTE
+};
+
+static inline constexpr bool hasShaderType(ShaderStageFlags flags, ShaderStage type) noexcept {
+    switch (type) {
+        case ShaderStage::VERTEX:
+            return bool(uint8_t(flags) & uint8_t(ShaderStageFlags::VERTEX));
+        case ShaderStage::FRAGMENT:
+            return bool(uint8_t(flags) & uint8_t(ShaderStageFlags::FRAGMENT));
+        case ShaderStage::COMPUTE:
+            return bool(uint8_t(flags) & uint8_t(ShaderStageFlags::COMPUTE));
+    }
+}
+
+enum class DescriptorType : uint8_t {
+    UNIFORM_BUFFER,
+    SHADER_STORAGE_BUFFER,
+    SAMPLER,
+    INPUT_ATTACHMENT,
+};
+
+enum class DescriptorFlags : uint8_t {
+    NONE = 0x00,
+    DYNAMIC_OFFSET = 0x01
+};
+
+using descriptor_set_t = uint8_t;
+
+using descriptor_binding_t = uint8_t;
+
+struct DescriptorSetLayoutBinding {
+    DescriptorType type;
+    ShaderStageFlags stageFlags;
+    descriptor_binding_t binding;
+    DescriptorFlags flags = DescriptorFlags::NONE;
+    uint16_t count = 0;
+
+    friend inline bool operator==(
+            DescriptorSetLayoutBinding const& lhs,
+            DescriptorSetLayoutBinding const& rhs) noexcept {
+        return lhs.type == rhs.type &&
+               lhs.flags == rhs.flags &&
+               lhs.count == rhs.count &&
+               lhs.stageFlags == rhs.stageFlags;
+    }
+};
+
+struct DescriptorSetLayout {
+    utils::FixedCapacityVector<DescriptorSetLayoutBinding> bindings;
+};
+
+
 /**
  * Bitmask for selecting render buffers
  */
@@ -270,15 +339,6 @@ enum class FenceStatus : int8_t {
     TIMEOUT_EXPIRED = 1,        //!< wait()'s timeout expired. The Fence condition is not satisfied.
 };
 
-/**
- * Status codes for sync objects
- */
-enum class SyncStatus : int8_t {
-    ERROR = -1,          //!< An error occurred. The Sync is not signaled.
-    SIGNALED = 0,        //!< The Sync is signaled.
-    NOT_SIGNALED = 1,    //!< The Sync is not signaled yet
-};
-
 static constexpr uint64_t FENCE_WAIT_FOR_EVER = uint64_t(-1);
 
 /**
@@ -368,6 +428,18 @@ enum class SamplerType : uint8_t {
     SAMPLER_CUBEMAP_ARRAY,  //!< Cube map array texture (feature level 2)
 };
 
+inline const char* stringify(SamplerType samplerType) {
+    switch (samplerType) {
+        case SamplerType::SAMPLER_2D: return "SAMPLER_2D";
+        case SamplerType::SAMPLER_2D_ARRAY: return "SAMPLER_2D_ARRAY";
+        case SamplerType::SAMPLER_CUBEMAP: return "SAMPLER_CUBEMAP";
+        case SamplerType::SAMPLER_EXTERNAL: return "SAMPLER_EXTERNAL";
+        case SamplerType::SAMPLER_3D: return "SAMPLER_3D";
+        case SamplerType::SAMPLER_CUBEMAP_ARRAY: return "SAMPLER_CUBEMAP_ARRAY";
+    }
+    return "UNKNOWN";
+}
+
 //! Subpass type
 enum class SubpassType : uint8_t {
     SUBPASS_INPUT
@@ -693,9 +765,27 @@ enum class TextureUsage : uint16_t {
     BLIT_SRC            = 0x0040,            //!< Texture can be used the source of a blit()
     BLIT_DST            = 0x0080,            //!< Texture can be used the destination of a blit()
     PROTECTED           = 0x0100,            //!< Texture can be used for protected content
-    DEFAULT             = UPLOADABLE | SAMPLEABLE   //!< Default texture usage
+    DEFAULT             = UPLOADABLE | SAMPLEABLE,   //!< Default texture usage
+    ALL_ATTACHMENTS     = COLOR_ATTACHMENT | DEPTH_ATTACHMENT | STENCIL_ATTACHMENT | SUBPASS_INPUT,   //!< Mask of all attachments
 };
 
+inline const char* stringify(TextureUsage usage) {
+    switch (usage) {
+        case TextureUsage::NONE: return "NONE";
+        case TextureUsage::COLOR_ATTACHMENT: return "COLOR_ATTACHMENT";
+        case TextureUsage::DEPTH_ATTACHMENT: return "DEPTH_ATTACHMENT";
+        case TextureUsage::STENCIL_ATTACHMENT: return "STENCIL_ATTACHMENT";
+        case TextureUsage::UPLOADABLE: return "UPLOADABLE";
+        case TextureUsage::SAMPLEABLE: return "SAMPLEABLE";
+        case TextureUsage::SUBPASS_INPUT: return "SUBPASS_INPUT";
+        case TextureUsage::BLIT_SRC: return "BLIT_SRC";
+        case TextureUsage::BLIT_DST: return "BLIT_DST";
+        case TextureUsage::PROTECTED: return "PROTECTED";
+        case TextureUsage::DEFAULT: return "DEFAULT";
+        default: return "UNKNOWN";
+    }
+}
+
 //! Texture swizzle
 enum class TextureSwizzle : uint8_t {
     SUBSTITUTE_ZERO,
@@ -887,6 +977,9 @@ struct SamplerParams { // NOLINT
 
     struct EqualTo {
         bool operator()(SamplerParams lhs, SamplerParams rhs) const noexcept {
+            assert_invariant(lhs.padding0 == 0);
+            assert_invariant(lhs.padding1 == 0);
+            assert_invariant(lhs.padding2 == 0);
             auto* pLhs = reinterpret_cast<uint32_t const*>(reinterpret_cast<char const*>(&lhs));
             auto* pRhs = reinterpret_cast<uint32_t const*>(reinterpret_cast<char const*>(&rhs));
             return *pLhs == *pRhs;
@@ -895,6 +988,9 @@ struct SamplerParams { // NOLINT
 
     struct LessThan {
         bool operator()(SamplerParams lhs, SamplerParams rhs) const noexcept {
+            assert_invariant(lhs.padding0 == 0);
+            assert_invariant(lhs.padding1 == 0);
+            assert_invariant(lhs.padding2 == 0);
             auto* pLhs = reinterpret_cast<uint32_t const*>(reinterpret_cast<char const*>(&lhs));
             auto* pRhs = reinterpret_cast<uint32_t const*>(reinterpret_cast<char const*>(&rhs));
             return *pLhs == *pRhs;
@@ -902,6 +998,12 @@ struct SamplerParams { // NOLINT
     };
 
 private:
+    friend inline bool operator == (SamplerParams lhs, SamplerParams rhs) noexcept {
+        return SamplerParams::EqualTo{}(lhs, rhs);
+    }
+    friend inline bool operator != (SamplerParams lhs, SamplerParams rhs) noexcept {
+        return  !SamplerParams::EqualTo{}(lhs, rhs);
+    }
     friend inline bool operator < (SamplerParams lhs, SamplerParams rhs) noexcept {
         return SamplerParams::LessThan{}(lhs, rhs);
     }
@@ -1058,7 +1160,7 @@ struct RasterState {
             bool inverseFrontFaces                      : 1;        // 31
 
             //! padding, must be 0
-            uint8_t padding                             : 1;        // 32
+            bool depthClamp                             : 1;        // 32
         };
         uint32_t u = 0;
     };
@@ -1069,32 +1171,6 @@ struct RasterState {
  * \privatesection
  */
 
-enum class ShaderStage : uint8_t {
-    VERTEX = 0,
-    FRAGMENT = 1,
-    COMPUTE = 2
-};
-
-static constexpr size_t PIPELINE_STAGE_COUNT = 2;
-enum class ShaderStageFlags : uint8_t {
-    NONE        =    0,
-    VERTEX      =    0x1,
-    FRAGMENT    =    0x2,
-    COMPUTE     =    0x4,
-    ALL_SHADER_STAGE_FLAGS = VERTEX | FRAGMENT | COMPUTE
-};
-
-static inline constexpr bool hasShaderType(ShaderStageFlags flags, ShaderStage type) noexcept {
-    switch (type) {
-        case ShaderStage::VERTEX:
-            return bool(uint8_t(flags) & uint8_t(ShaderStageFlags::VERTEX));
-        case ShaderStage::FRAGMENT:
-            return bool(uint8_t(flags) & uint8_t(ShaderStageFlags::FRAGMENT));
-        case ShaderStage::COMPUTE:
-            return bool(uint8_t(flags) & uint8_t(ShaderStageFlags::COMPUTE));
-    }
-}
-
 /**
  * Selects which buffers to clear at the beginning of the render pass, as well as which buffers
  * can be discarded at the beginning and end of the render pass.
@@ -1244,7 +1320,7 @@ enum class Workaround : uint16_t {
     ADRENO_UNIFORM_ARRAY_CRASH,
     // Workaround a Metal pipeline compilation error with the message:
     // "Could not statically determine the target of a texture". See light_indirect.fs
-    A8X_STATIC_TEXTURE_TARGET_ERROR,
+    METAL_STATIC_TEXTURE_TARGET_ERROR,
     // Adreno drivers sometimes aren't able to blit into a layer of a texture array.
     DISABLE_BLIT_INTO_TEXTURE_ARRAY,
     // Multiple workarounds needed for PowerVR GPUs
@@ -1259,6 +1335,8 @@ template<> struct utils::EnableBitMaskOperators<filament::backend::ShaderStageFl
         : public std::true_type {};
 template<> struct utils::EnableBitMaskOperators<filament::backend::TargetBufferFlags>
         : public std::true_type {};
+template<> struct utils::EnableBitMaskOperators<filament::backend::DescriptorFlags>
+        : public std::true_type {};
 template<> struct utils::EnableBitMaskOperators<filament::backend::TextureUsage>
         : public std::true_type {};
 template<> struct utils::EnableBitMaskOperators<filament::backend::StencilFace>

filament/backend/include/backend/Handle.h

@@ -23,6 +23,7 @@
 #include <utils/debug.h>
 
 #include <type_traits> // FIXME: STL headers are not allowed in public headers
+#include <utility>
 
 #include <stdint.h>
 
@@ -41,6 +42,8 @@ struct HwTexture;
 struct HwTimerQuery;
 struct HwVertexBufferInfo;
 struct HwVertexBuffer;
+struct HwDescriptorSetLayout;
+struct HwDescriptorSet;
 
 /*
  * A handle to a backend resource. HandleBase is for internal use only.
@@ -104,8 +107,18 @@ struct Handle : public HandleBase {
     Handle(Handle const& rhs) noexcept = default;
     Handle(Handle&& rhs) noexcept = default;
 
-    Handle& operator=(Handle const& rhs) noexcept = default;
-    Handle& operator=(Handle&& rhs) noexcept = default;
+    // Explicitly redefine copy/move assignment operators rather than just using default here.
+    // Because it doesn't make a call to the parent's method automatically during the std::move
+    // function call(https://en.cppreference.com/w/cpp/algorithm/move) in certain compilers like
+    // NDK 25.1.8937393 and below (see b/371980551)
+    Handle& operator=(Handle const& rhs) noexcept {
+        HandleBase::operator=(rhs);
+        return *this;
+    }
+    Handle& operator=(Handle&& rhs) noexcept {
+        HandleBase::operator=(std::move(rhs));
+        return *this;
+    }
 
     explicit Handle(HandleId id) noexcept : HandleBase(id) { }
 
@@ -130,19 +143,21 @@ private:
 
 // Types used by the command stream
 // (we use this renaming because the macro-system doesn't deal well with "<" and ">")
-using BufferObjectHandle     = Handle<HwBufferObject>;
-using FenceHandle            = Handle<HwFence>;
-using IndexBufferHandle      = Handle<HwIndexBuffer>;
-using ProgramHandle          = Handle<HwProgram>;
-using RenderPrimitiveHandle  = Handle<HwRenderPrimitive>;
-using RenderTargetHandle     = Handle<HwRenderTarget>;
-using SamplerGroupHandle     = Handle<HwSamplerGroup>;
-using StreamHandle           = Handle<HwStream>;
-using SwapChainHandle        = Handle<HwSwapChain>;
-using TextureHandle          = Handle<HwTexture>;
-using TimerQueryHandle       = Handle<HwTimerQuery>;
-using VertexBufferHandle     = Handle<HwVertexBuffer>;
-using VertexBufferInfoHandle = Handle<HwVertexBufferInfo>;
+using BufferObjectHandle        = Handle<HwBufferObject>;
+using FenceHandle               = Handle<HwFence>;
+using IndexBufferHandle         = Handle<HwIndexBuffer>;
+using ProgramHandle             = Handle<HwProgram>;
+using RenderPrimitiveHandle     = Handle<HwRenderPrimitive>;
+using RenderTargetHandle        = Handle<HwRenderTarget>;
+using SamplerGroupHandle        = Handle<HwSamplerGroup>;
+using StreamHandle              = Handle<HwStream>;
+using SwapChainHandle           = Handle<HwSwapChain>;
+using TextureHandle             = Handle<HwTexture>;
+using TimerQueryHandle          = Handle<HwTimerQuery>;
+using VertexBufferHandle        = Handle<HwVertexBuffer>;
+using VertexBufferInfoHandle    = Handle<HwVertexBufferInfo>;
+using DescriptorSetLayoutHandle = Handle<HwDescriptorSetLayout>;
+using DescriptorSetHandle       = Handle<HwDescriptorSet>;
 
 } // namespace filament::backend
 

filament/backend/include/backend/PipelineState.h

@@ -22,15 +22,23 @@
 
 #include <utils/ostream.h>
 
+#include <array>
+
 #include <stdint.h>
 
 namespace filament::backend {
 
 //! \privatesection
 
+struct PipelineLayout {
+    using SetLayout = std::array<Handle<HwDescriptorSetLayout>, MAX_DESCRIPTOR_SET_COUNT>;
+    SetLayout setLayout;      // 16
+};
+
 struct PipelineState {
     Handle<HwProgram> program;                                              //  4
     Handle<HwVertexBufferInfo> vertexBufferInfo;                            //  4
+    PipelineLayout pipelineLayout;                                          // 16
     RasterState rasterState;                                                //  4
     StencilState stencilState;                                              // 12
     PolygonOffset polygonOffset;                                            //  8

filament/backend/include/backend/Platform.h

@@ -68,12 +68,7 @@ public:
          */
         size_t handleArenaSize = 0;
 
-        /**
-         * This number of most-recently destroyed textures will be tracked for use-after-free.
-         * Throws an exception when a texture is freed but still bound to a SamplerGroup and used in
-         * a draw call. 0 disables completely. Currently only respected by the Metal backend.
-         */
-        size_t textureUseAfterFreePoolSize = 0;
+        size_t metalUploadBufferSizeBytes = 512 * 1024;
 
         /**
          * Set to `true` to forcibly disable parallel shader compilation in the backend.
@@ -96,6 +91,13 @@ public:
          * Sets the technique for stereoscopic rendering.
          */
         StereoscopicType stereoscopicType = StereoscopicType::NONE;
+
+        /**
+         * Assert the native window associated to a SwapChain is valid when calling makeCurrent().
+         * This is only supported for:
+         *      - PlatformEGLAndroid
+         */
+        bool assertNativeWindowIsValid = false;
     };
 
     Platform() noexcept;

filament/backend/include/backend/Program.h

@@ -24,9 +24,11 @@
 
 #include <backend/DriverEnums.h>
 
-#include <array>    // FIXME: STL headers are not allowed in public headers
-#include <utility>  // FIXME: STL headers are not allowed in public headers
-#include <variant>  // FIXME: STL headers are not allowed in public headers
+#include <array>
+#include <unordered_map>
+#include <tuple>
+#include <utility>
+#include <variant>
 
 #include <stddef.h>
 #include <stdint.h>
@@ -40,29 +42,36 @@ public:
     static constexpr size_t UNIFORM_BINDING_COUNT = CONFIG_UNIFORM_BINDING_COUNT;
     static constexpr size_t SAMPLER_BINDING_COUNT = CONFIG_SAMPLER_BINDING_COUNT;
 
-    struct Sampler {
-        utils::CString name = {};   // name of the sampler in the shader
-        uint32_t binding = 0;       // binding point of the sampler in the shader
+    struct Descriptor {
+        utils::CString name;
+        backend::DescriptorType type;
+        backend::descriptor_binding_t binding;
     };
 
-    struct SamplerGroupData {
-        utils::FixedCapacityVector<Sampler> samplers;
-        ShaderStageFlags stageFlags = ShaderStageFlags::ALL_SHADER_STAGE_FLAGS;
+    struct SpecializationConstant {
+        using Type = std::variant<int32_t, float, bool>;
+        uint32_t id;    // id set in glsl
+        Type value;     // value and type
     };
 
-    struct Uniform {
+    struct Uniform { // For ES2 support
         utils::CString name;    // full qualified name of the uniform field
         uint16_t offset;        // offset in 'uint32_t' into the uniform buffer
         uint8_t size;           // >1 for arrays
         UniformType type;       // uniform type
     };
 
-    using UniformBlockInfo = std::array<utils::CString, UNIFORM_BINDING_COUNT>;
-    using UniformInfo = utils::FixedCapacityVector<Uniform>;
-    using SamplerGroupInfo = std::array<SamplerGroupData, SAMPLER_BINDING_COUNT>;
+    using DescriptorBindingsInfo = utils::FixedCapacityVector<Descriptor>;
+    using DescriptorSetInfo = std::array<DescriptorBindingsInfo, MAX_DESCRIPTOR_SET_COUNT>;
+    using SpecializationConstantsInfo = utils::FixedCapacityVector<SpecializationConstant>;
     using ShaderBlob = utils::FixedCapacityVector<uint8_t>;
     using ShaderSource = std::array<ShaderBlob, SHADER_TYPE_COUNT>;
 
+    using AttributesInfo = utils::FixedCapacityVector<std::pair<utils::CString, uint8_t>>;
+    using UniformInfo = utils::FixedCapacityVector<Uniform>;
+    using BindingUniformsInfo = utils::FixedCapacityVector<
+            std::tuple<uint8_t, utils::CString, Program::UniformInfo>>;
+
     Program() noexcept;
 
     Program(const Program& rhs) = delete;
@@ -79,43 +88,19 @@ public:
     Program& diagnostics(utils::CString const& name,
             utils::Invocable<utils::io::ostream&(utils::io::ostream& out)>&& logger);
 
-    // sets one of the program's shader (e.g. vertex, fragment)
+    // Sets one of the program's shader (e.g. vertex, fragment)
     // string-based shaders are null terminated, consequently the size parameter must include the
     // null terminating character.
     Program& shader(ShaderStage shader, void const* data, size_t size);
 
-    // sets the language of the shader sources provided with shader() (defaults to ESSL3)
+    // Sets the language of the shader sources provided with shader() (defaults to ESSL3)
     Program& shaderLanguage(ShaderLanguage shaderLanguage);
 
-    // Note: This is only needed for GLES3.0 backends, because the layout(binding=) syntax is
-    //       not permitted in glsl. The backend needs a way to associate a uniform block
-    //       to a binding point.
-    Program& uniformBlockBindings(
-            utils::FixedCapacityVector<std::pair<utils::CString, uint8_t>> const& uniformBlockBindings) noexcept;
+    // Descriptor binding (set, binding, type -> shader name) info
+    Program& descriptorBindings(backend::descriptor_set_t set,
+            DescriptorBindingsInfo descriptorBindings) noexcept;
 
-    // Note: This is only needed for GLES2.0, this is used to emulate UBO. This function tells
-    // the program everything it needs to know about the uniforms at a given binding
-    Program& uniforms(uint32_t index, UniformInfo const& uniforms) noexcept;
-
-    // Note: This is only needed for GLES2.0.
-    Program& attributes(
-            utils::FixedCapacityVector<std::pair<utils::CString, uint8_t>> attributes) noexcept;
-
-    // sets the 'bindingPoint' sampler group descriptor for this program.
-    // 'samplers' can be destroyed after this call.
-    // This effectively associates a set of (BindingPoints, index) to a texture unit in the shader.
-    // Or more precisely, what layout(binding=) is set to in GLSL.
-    Program& setSamplerGroup(size_t bindingPoint, ShaderStageFlags stageFlags,
-            Sampler const* samplers, size_t count) noexcept;
-
-    struct SpecializationConstant {
-        using Type = std::variant<int32_t, float, bool>;
-        uint32_t id;                                // id set in glsl
-        Type value;                                 // value and type
-    };
-
-    Program& specializationConstants(
-            utils::FixedCapacityVector<SpecializationConstant> specConstants) noexcept;
+    Program& specializationConstants(SpecializationConstantsInfo specConstants) noexcept;
 
     struct PushConstant {
         utils::CString name;
@@ -129,33 +114,40 @@ public:
 
     Program& multiview(bool multiview) noexcept;
 
+    // For ES2 support only...
+    Program& uniforms(uint32_t index, utils::CString name, UniformInfo uniforms) noexcept;
+    Program& attributes(AttributesInfo attributes) noexcept;
+
+    //
+    // Getters for program construction...
+    //
+
     ShaderSource const& getShadersSource() const noexcept { return mShadersSource; }
     ShaderSource& getShadersSource() noexcept { return mShadersSource; }
 
-    UniformBlockInfo const& getUniformBlockBindings() const noexcept { return mUniformBlocks; }
-    UniformBlockInfo& getUniformBlockBindings() noexcept { return mUniformBlocks; }
-
-    SamplerGroupInfo const& getSamplerGroupInfo() const { return mSamplerGroups; }
-    SamplerGroupInfo& getSamplerGroupInfo() { return mSamplerGroups; }
-
-    auto const& getBindingUniformInfo() const { return mBindingUniformInfo; }
-    auto& getBindingUniformInfo() { return mBindingUniformInfo; }
-
-    auto const& getAttributes() const { return mAttributes; }
-    auto& getAttributes() { return mAttributes; }
-
     utils::CString const& getName() const noexcept { return mName; }
     utils::CString& getName() noexcept { return mName; }
 
     auto const& getShaderLanguage() const { return mShaderLanguage; }
 
-    utils::FixedCapacityVector<SpecializationConstant> const& getSpecializationConstants() const noexcept {
+    uint64_t getCacheId() const noexcept { return mCacheId; }
+
+    bool isMultiview() const noexcept { return mMultiview; }
+
+    CompilerPriorityQueue getPriorityQueue() const noexcept { return mPriorityQueue; }
+
+    SpecializationConstantsInfo const& getSpecializationConstants() const noexcept {
         return mSpecializationConstants;
     }
-    utils::FixedCapacityVector<SpecializationConstant>& getSpecializationConstants() noexcept {
+
+    SpecializationConstantsInfo& getSpecializationConstants() noexcept {
         return mSpecializationConstants;
     }
 
+    DescriptorSetInfo& getDescriptorBindings() noexcept {
+        return mDescriptorBindings;
+    }
+
     utils::FixedCapacityVector<PushConstant> const& getPushConstants(
             ShaderStage stage) const noexcept {
         return mPushConstants[static_cast<uint8_t>(stage)];
@@ -165,27 +157,29 @@ public:
         return mPushConstants[static_cast<uint8_t>(stage)];
     }
 
-    uint64_t getCacheId() const noexcept { return mCacheId; }
+    auto const& getBindingUniformInfo() const { return mBindingUniformsInfo; }
+    auto& getBindingUniformInfo() { return mBindingUniformsInfo; }
 
-    bool isMultiview() const noexcept { return mMultiview; }
-
-    CompilerPriorityQueue getPriorityQueue() const noexcept { return mPriorityQueue; }
+    auto const& getAttributes() const { return mAttributes; }
+    auto& getAttributes() { return mAttributes; }
 
 private:
     friend utils::io::ostream& operator<<(utils::io::ostream& out, const Program& builder);
 
-    UniformBlockInfo mUniformBlocks = {};
-    SamplerGroupInfo mSamplerGroups = {};
     ShaderSource mShadersSource;
     ShaderLanguage mShaderLanguage = ShaderLanguage::ESSL3;
     utils::CString mName;
     uint64_t mCacheId{};
-    utils::Invocable<utils::io::ostream&(utils::io::ostream& out)> mLogger;
-    utils::FixedCapacityVector<SpecializationConstant> mSpecializationConstants;
-    std::array<utils::FixedCapacityVector<PushConstant>, SHADER_TYPE_COUNT> mPushConstants;
-    utils::FixedCapacityVector<std::pair<utils::CString, uint8_t>> mAttributes;
-    std::array<UniformInfo, Program::UNIFORM_BINDING_COUNT> mBindingUniformInfo;
     CompilerPriorityQueue mPriorityQueue = CompilerPriorityQueue::HIGH;
+    utils::Invocable<utils::io::ostream&(utils::io::ostream& out)> mLogger;
+    SpecializationConstantsInfo mSpecializationConstants;
+    std::array<utils::FixedCapacityVector<PushConstant>, SHADER_TYPE_COUNT> mPushConstants;
+    DescriptorSetInfo mDescriptorBindings;
+
+    // For ES2 support only
+    AttributesInfo mAttributes;
+    BindingUniformsInfo mBindingUniformsInfo;
+
     // Indicates the current engine was initialized with multiview stereo, and the variant for this
     // program contains STE flag. This will be referred later for the OpenGL shader compiler to
     // determine whether shader code replacement for the num_views should be performed.

filament/backend/include/backend/TargetBufferInfo.h

@@ -30,10 +30,6 @@ namespace filament::backend {
 
 struct TargetBufferInfo {
     // note: the parameters of this constructor are not in the order of this structure's fields
-    TargetBufferInfo(Handle<HwTexture> handle, uint8_t level, uint16_t layer, uint8_t baseViewIndex) noexcept
-        : handle(handle), baseViewIndex(baseViewIndex), level(level), layer(layer) {
-    }
-
     TargetBufferInfo(Handle<HwTexture> handle, uint8_t level, uint16_t layer) noexcept
             : handle(handle), level(level), layer(layer) {
     }
@@ -51,14 +47,15 @@ struct TargetBufferInfo {
     // texture to be used as render target
     Handle<HwTexture> handle;
 
-    // Starting layer index for multiview. This value is only used when the `layerCount` for the
-    // render target is greater than 1.
-    uint8_t baseViewIndex = 0;
-
     // level to be used
     uint8_t level = 0;
 
-    // For cubemaps and 3D textures. See TextureCubemapFace for the face->layer mapping
+    // - For cubemap textures, this indicates the face of the cubemap. See TextureCubemapFace for
+    //   the face->layer mapping)
+    // - For 2d array, cubemap array, and 3d textures, this indicates an index of a single layer of
+    //   them.
+    // - For multiview textures (i.e., layerCount for the RenderTarget is greater than 1), this
+    //   indicates a starting layer index of the current 2d array texture for multiview.
     uint16_t layer = 0;
 };
 
@@ -103,7 +100,7 @@ public:
 
     // this is here for backward compatibility
     MRT(Handle<HwTexture> handle, uint8_t level, uint16_t layer) noexcept
-            : mInfos{{ handle, level, layer, 0 }} {
+            : mInfos{{ handle, level, layer }} {
     }
 };
 

filament/backend/include/backend/platforms/PlatformEGLAndroid.h

@@ -89,14 +89,28 @@ protected:
      */
     AcquiredImage transformAcquiredImage(AcquiredImage source) noexcept override;
 
+protected:
+    bool makeCurrent(ContextType type,
+            SwapChain* drawSwapChain,
+            SwapChain* readSwapChain) noexcept override;
+
 private:
+    struct InitializeJvmForPerformanceManagerIfNeeded {
+        InitializeJvmForPerformanceManagerIfNeeded();
+    };
+
     int mOSVersion;
     ExternalStreamManagerAndroid& mExternalStreamManager;
+    InitializeJvmForPerformanceManagerIfNeeded const mInitializeJvmForPerformanceManagerIfNeeded;
     utils::PerformanceHintManager mPerformanceHintManager;
     utils::PerformanceHintManager::Session mPerformanceHintSession;
 
     using clock = std::chrono::high_resolution_clock;
     clock::time_point mStartTimeOfActualWork;
+
+    void* mNativeWindowLib = nullptr;
+    int32_t (*ANativeWindow_getBuffersDefaultDataSpace)(ANativeWindow* window) = nullptr;
+    bool mAssertNativeWindowIsValid = false;
 };
 
 } // namespace filament::backend

filament/backend/include/backend/platforms/PlatformOSMesa.h

@@ -0,0 +1,64 @@
+/*
+ * Copyright (C) 2024 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef TNT_FILAMENT_BACKEND_OPENGL_OPENGL_PLATFORM_OSMESA_H
+#define TNT_FILAMENT_BACKEND_OPENGL_OPENGL_PLATFORM_OSMESA_H
+
+#include <stdint.h>
+
+#include "bluegl/BlueGL.h"
+
+#include <osmesa.h>
+
+#include <backend/platforms/OpenGLPlatform.h>
+#include <backend/DriverEnums.h>
+
+namespace filament::backend {
+
+/**
+ * A concrete implementation of OpenGLPlatform that uses OSMesa, which is an offscreen
+ * context that can be used in conjunction with Mesa for software rasterization.
+ * See https://docs.mesa3d.org/osmesa.html for more information.
+ */
+class PlatformOSMesa : public OpenGLPlatform {
+protected:
+    // --------------------------------------------------------------------------------------------
+    // Platform Interface
+
+    Driver* createDriver(void* sharedGLContext, const DriverConfig& driverConfig) noexcept override;
+
+    int getOSVersion() const noexcept final override { return 0; }
+
+    // --------------------------------------------------------------------------------------------
+    // OpenGLPlatform Interface
+
+    void terminate() noexcept override;
+
+    SwapChain* createSwapChain(void* nativewindow, uint64_t flags) noexcept override;
+    SwapChain* createSwapChain(uint32_t width, uint32_t height, uint64_t flags) noexcept override;
+    void destroySwapChain(SwapChain* swapChain) noexcept override;
+    bool makeCurrent(ContextType type, SwapChain* drawSwapChain,
+            SwapChain* readSwapChain) noexcept override;
+    void commit(SwapChain* swapChain) noexcept override;
+
+private:
+    OSMesaContext mContext;
+    void* mOsMesaApi = nullptr;
+};
+
+} // namespace filament::backend
+
+#endif // TNT_FILAMENT_BACKEND_OPENGL_OPENGL_PLATFORM_OSMESA_H

filament/backend/include/backend/platforms/VulkanPlatform.h

@@ -90,6 +90,20 @@ public:
         VkExtent2D extent = {0, 0};
     };
 
+    struct ImageSyncData {
+        static constexpr uint32_t INVALID_IMAGE_INDEX = UINT32_MAX;
+
+        // The index of the next image as returned by vkAcquireNextImage or equivalent.
+        uint32_t imageIndex = INVALID_IMAGE_INDEX;
+
+        // Semaphore to be signaled once the image is available.
+        VkSemaphore imageReadySemaphore = VK_NULL_HANDLE;
+
+        // A function called right before vkQueueSubmit. After this call, the image must be 
+        // available. This pointer can be null if imageReadySemaphore is not VK_NULL_HANDLE.
+        std::function<void(SwapChainPtr handle)> explicitImageReadyWait = nullptr;
+    };
+
     VulkanPlatform();
 
     ~VulkanPlatform() override;
@@ -127,6 +141,12 @@ public:
          * before recreating the swapchain. Default is true.
          */
         bool flushAndWaitOnWindowResize = true;
+
+        /**
+         * Whether the swapchain image should be transitioned to a layout suitable for
+         * presentation. Default is true.
+         */
+        bool transitionSwapChainImageLayoutForPresent = true;
     };
 
     /**
@@ -155,13 +175,10 @@ public:
      * corresponding VkImage will be used as the output color attachment. The client should signal
      * the `clientSignal` semaphore when the image is ready to be used by the backend.
      * @param handle         The handle returned by createSwapChain()
-     * @param clientSignal   The semaphore that the client will signal to indicate that the backend
-     *                       may render into the image.
-     * @param index          Pointer to memory that will be filled with the index that corresponding
-     *                       to an image in the `SwapChainBundle.colors` array.
+     * @param outImageSyncData The synchronization data used for image readiness
      * @return               Result of acquire
      */
-    virtual VkResult acquire(SwapChainPtr handle, VkSemaphore clientSignal, uint32_t* index);
+    virtual VkResult acquire(SwapChainPtr handle, ImageSyncData* outImageSyncData);
 
     /**
      * Present the image corresponding to `index` to the display. The client should wait on

filament/backend/include/private/backend/Driver.h

@@ -18,6 +18,7 @@
 #define TNT_FILAMENT_BACKEND_PRIVATE_DRIVER_H
 
 #include <backend/CallbackHandler.h>
+#include <backend/DescriptorSetOffsetArray.h>
 #include <backend/DriverApiForward.h>
 #include <backend/DriverEnums.h>
 #include <backend/Handle.h>

filament/backend/include/private/backend/DriverAPI.inc

@@ -139,7 +139,8 @@ DECL_DRIVER_API_N(beginFrame,
 DECL_DRIVER_API_N(setFrameScheduledCallback,
         backend::SwapChainHandle, sch,
         backend::CallbackHandler*, handler,
-        backend::FrameScheduledCallback&&, callback)
+        backend::FrameScheduledCallback&&, callback,
+        uint64_t, flags)
 
 DECL_DRIVER_API_N(setFrameCompletedCallback,
         backend::SwapChainHandle, sch,
@@ -162,6 +163,10 @@ DECL_DRIVER_API_0(finish)
 // reset state tracking, if the driver does any state tracking (e.g. GL)
 DECL_DRIVER_API_0(resetState)
 
+DECL_DRIVER_API_N(setDebugTag,
+        backend::HandleBase::HandleId, handleId,
+        utils::CString, tag)
+
 /*
  * Creating driver objects
  * -----------------------
@@ -196,20 +201,33 @@ DECL_DRIVER_API_R_N(backend::TextureHandle, createTexture,
         uint32_t, depth,
         backend::TextureUsage, usage)
 
-DECL_DRIVER_API_R_N(backend::TextureHandle, createTextureSwizzled,
-        backend::SamplerType, target,
-        uint8_t, levels,
-        backend::TextureFormat, format,
-        uint8_t, samples,
-        uint32_t, width,
-        uint32_t, height,
-        uint32_t, depth,
-        backend::TextureUsage, usage,
+DECL_DRIVER_API_R_N(backend::TextureHandle, createTextureView,
+        backend::TextureHandle, texture,
+        uint8_t, baseLevel,
+        uint8_t, levelCount)
+
+DECL_DRIVER_API_R_N(backend::TextureHandle, createTextureViewSwizzle,
+        backend::TextureHandle, texture,
         backend::TextureSwizzle, r,
         backend::TextureSwizzle, g,
         backend::TextureSwizzle, b,
         backend::TextureSwizzle, a)
 
+DECL_DRIVER_API_R_N(backend::TextureHandle, createTextureExternalImage,
+        backend::TextureFormat, format,
+        uint32_t, width,
+        uint32_t, height,
+        backend::TextureUsage, usage,
+        void*, image)
+
+DECL_DRIVER_API_R_N(backend::TextureHandle, createTextureExternalImagePlane,
+        backend::TextureFormat, format,
+        uint32_t, width,
+        uint32_t, height,
+        backend::TextureUsage, usage,
+        void*, image,
+        uint32_t, plane)
+
 DECL_DRIVER_API_R_N(backend::TextureHandle, importTexture,
         intptr_t, id,
         backend::SamplerType, target,
@@ -221,9 +239,6 @@ DECL_DRIVER_API_R_N(backend::TextureHandle, importTexture,
         uint32_t, depth,
         backend::TextureUsage, usage)
 
-DECL_DRIVER_API_R_N(backend::SamplerGroupHandle, createSamplerGroup,
-        uint32_t, size, utils::FixedSizeString<32>, debugName)
-
 DECL_DRIVER_API_R_N(backend::RenderPrimitiveHandle, createRenderPrimitive,
         backend::VertexBufferHandle, vbh,
         backend::IndexBufferHandle, ibh,
@@ -257,25 +272,53 @@ DECL_DRIVER_API_R_N(backend::SwapChainHandle, createSwapChainHeadless,
 
 DECL_DRIVER_API_R_0(backend::TimerQueryHandle, createTimerQuery)
 
+DECL_DRIVER_API_R_N(backend::DescriptorSetLayoutHandle, createDescriptorSetLayout,
+        backend::DescriptorSetLayout&&, info)
+
+DECL_DRIVER_API_R_N(backend::DescriptorSetHandle, createDescriptorSet,
+        backend::DescriptorSetLayoutHandle, dslh)
+
+DECL_DRIVER_API_N(updateDescriptorSetBuffer,
+        backend::DescriptorSetHandle, dsh,
+        backend::descriptor_binding_t, binding,
+        backend::BufferObjectHandle, boh,
+        uint32_t, offset,
+        uint32_t, size
+)
+
+DECL_DRIVER_API_N(updateDescriptorSetTexture,
+        backend::DescriptorSetHandle, dsh,
+        backend::descriptor_binding_t, binding,
+        backend::TextureHandle, th,
+        SamplerParams, params
+)
+
+DECL_DRIVER_API_N(bindDescriptorSet,
+        backend::DescriptorSetHandle, dsh,
+        backend::descriptor_set_t, set,
+        backend::DescriptorSetOffsetArray&&, offsets
+)
+
 
 /*
  * Destroying driver objects
  * -------------------------
  */
 
-DECL_DRIVER_API_N(destroyVertexBuffer,    backend::VertexBufferHandle, vbh)
-DECL_DRIVER_API_N(destroyVertexBufferInfo,backend::VertexBufferInfoHandle, vbih)
-DECL_DRIVER_API_N(destroyIndexBuffer,     backend::IndexBufferHandle, ibh)
-DECL_DRIVER_API_N(destroyBufferObject,    backend::BufferObjectHandle, ibh)
-DECL_DRIVER_API_N(destroyRenderPrimitive, backend::RenderPrimitiveHandle, rph)
-DECL_DRIVER_API_N(destroyProgram,         backend::ProgramHandle, ph)
-DECL_DRIVER_API_N(destroySamplerGroup,    backend::SamplerGroupHandle, sbh)
-DECL_DRIVER_API_N(destroyTexture,         backend::TextureHandle, th)
-DECL_DRIVER_API_N(destroyRenderTarget,    backend::RenderTargetHandle, rth)
-DECL_DRIVER_API_N(destroySwapChain,       backend::SwapChainHandle, sch)
-DECL_DRIVER_API_N(destroyStream,          backend::StreamHandle, sh)
-DECL_DRIVER_API_N(destroyTimerQuery,      backend::TimerQueryHandle, sh)
-DECL_DRIVER_API_N(destroyFence,           backend::FenceHandle, fh)
+DECL_DRIVER_API_N(destroyVertexBuffer,          backend::VertexBufferHandle, vbh)
+DECL_DRIVER_API_N(destroyVertexBufferInfo,      backend::VertexBufferInfoHandle, vbih)
+DECL_DRIVER_API_N(destroyIndexBuffer,           backend::IndexBufferHandle, ibh)
+DECL_DRIVER_API_N(destroyBufferObject,          backend::BufferObjectHandle, ibh)
+DECL_DRIVER_API_N(destroyRenderPrimitive,       backend::RenderPrimitiveHandle, rph)
+DECL_DRIVER_API_N(destroyProgram,               backend::ProgramHandle, ph)
+DECL_DRIVER_API_N(destroyTexture,               backend::TextureHandle, th)
+DECL_DRIVER_API_N(destroyRenderTarget,          backend::RenderTargetHandle, rth)
+DECL_DRIVER_API_N(destroySwapChain,             backend::SwapChainHandle, sch)
+DECL_DRIVER_API_N(destroyStream,                backend::StreamHandle, sh)
+DECL_DRIVER_API_N(destroyTimerQuery,            backend::TimerQueryHandle, sh)
+DECL_DRIVER_API_N(destroyFence,                 backend::FenceHandle, fh)
+DECL_DRIVER_API_N(destroyDescriptorSetLayout,   backend::DescriptorSetLayoutHandle, dslh)
+DECL_DRIVER_API_N(destroyDescriptorSet,         backend::DescriptorSetHandle, dsh)
 
 /*
  * Synchronous APIs
@@ -305,6 +348,7 @@ DECL_DRIVER_API_SYNCHRONOUS_0(bool, isParallelShaderCompileSupported)
 DECL_DRIVER_API_SYNCHRONOUS_0(bool, isDepthStencilResolveSupported)
 DECL_DRIVER_API_SYNCHRONOUS_N(bool, isDepthStencilBlitSupported, backend::TextureFormat, format)
 DECL_DRIVER_API_SYNCHRONOUS_0(bool, isProtectedTexturesSupported)
+DECL_DRIVER_API_SYNCHRONOUS_0(bool, isDepthClampSupported)
 DECL_DRIVER_API_SYNCHRONOUS_0(uint8_t, getMaxDrawBuffers)
 DECL_DRIVER_API_SYNCHRONOUS_0(size_t, getMaxUniformBufferSize)
 DECL_DRIVER_API_SYNCHRONOUS_0(math::float2, getClipSpaceParams)
@@ -341,15 +385,6 @@ DECL_DRIVER_API_N(updateBufferObjectUnsynchronized,
 DECL_DRIVER_API_N(resetBufferObject,
         backend::BufferObjectHandle, ibh)
 
-DECL_DRIVER_API_N(updateSamplerGroup,
-        backend::SamplerGroupHandle, ubh,
-        backend::BufferDescriptor&&, data)
-
-DECL_DRIVER_API_N(setMinMaxLevels,
-        backend::TextureHandle, th,
-        uint32_t, minLevel,
-        uint32_t, maxLevel)
-
 DECL_DRIVER_API_N(update3DImage,
         backend::TextureHandle, th,
         uint32_t, level,
@@ -364,10 +399,12 @@ DECL_DRIVER_API_N(update3DImage,
 DECL_DRIVER_API_N(generateMipmaps,
         backend::TextureHandle, th)
 
+// Deprecated
 DECL_DRIVER_API_N(setExternalImage,
         backend::TextureHandle, th,
         void*, image)
 
+// Deprecated
 DECL_DRIVER_API_N(setExternalImagePlane,
         backend::TextureHandle, th,
         void*, image,
@@ -414,37 +451,16 @@ DECL_DRIVER_API_N(commit,
  * -----------------------
  */
 
-DECL_DRIVER_API_N(bindUniformBuffer,
-        uint32_t, index,
-        backend::BufferObjectHandle, ubh)
-
-DECL_DRIVER_API_N(bindBufferRange,
-        BufferObjectBinding, bindingType,
-        uint32_t, index,
-        backend::BufferObjectHandle, ubh,
-        uint32_t, offset,
-        uint32_t, size)
-
-DECL_DRIVER_API_N(unbindBuffer,
-        BufferObjectBinding, bindingType,
-        uint32_t, index)
-
-DECL_DRIVER_API_N(bindSamplers,
-        uint32_t, index,
-        backend::SamplerGroupHandle, sbh)
-
 DECL_DRIVER_API_N(setPushConstant,
         backend::ShaderStage, stage,
         uint8_t, index,
         backend::PushConstantVariant, value)
 
 DECL_DRIVER_API_N(insertEventMarker,
-        const char*, string,
-        uint32_t, len = 0)
+        const char*, string)
 
 DECL_DRIVER_API_N(pushGroupMarker,
-        const char*, string,
-        uint32_t, len = 0)
+        const char*, string)
 
 DECL_DRIVER_API_0(popGroupMarker)
 

filament/backend/include/private/backend/DriverApi.h

@@ -18,6 +18,17 @@
 #define TNT_FILAMENT_BACKEND_PRIVATE_DRIVERAPI_H
 
 #include "backend/DriverApiForward.h"
+
 #include "private/backend/CommandStream.h"
 
+#include <stddef.h>
+
+namespace filament::backend {
+
+inline void* allocateFromCommandStream(DriverApi& driver, size_t size, size_t alignment) noexcept {
+    return driver.allocate(size, alignment);
+}
+
+} // namespace filament::backend
+
 #endif // TNT_FILAMENT_BACKEND_PRIVATE_DRIVERAPI_H

filament/backend/include/private/backend/HandleAllocator.h

@@ -20,11 +20,12 @@
 #include <backend/Handle.h>
 
 #include <utils/Allocator.h>
+#include <utils/CString.h>
 #include <utils/Log.h>
+#include <utils/Panic.h>
 #include <utils/compiler.h>
 #include <utils/debug.h>
 #include <utils/ostream.h>
-#include <utils/Panic.h>
 
 #include <tsl/robin_map.h>
 
@@ -37,9 +38,9 @@
 #include <stddef.h>
 #include <stdint.h>
 
-#define HandleAllocatorGL  HandleAllocator<32,  64, 136>    // ~4520 / pool / MiB
+#define HandleAllocatorGL  HandleAllocator<32,  96, 136>    // ~4520 / pool / MiB
 #define HandleAllocatorVK  HandleAllocator<64, 160, 312>    // ~1820 / pool / MiB
-#define HandleAllocatorMTL HandleAllocator<32,  48, 552>    // ~1660 / pool / MiB
+#define HandleAllocatorMTL HandleAllocator<32,  64, 552>    // ~1660 / pool / MiB
 
 namespace filament::backend {
 
@@ -168,19 +169,53 @@ public:
         auto [p, tag] = handleToPointer(handle.getId());
 
         if (isPoolHandle(handle.getId())) {
-            // check for use after free
+            // check for pool handle use-after-free
             if (UTILS_UNLIKELY(!mUseAfterFreeCheckDisabled)) {
                 uint8_t const age = (tag & HANDLE_AGE_MASK) >> HANDLE_AGE_SHIFT;
                 auto const pNode = static_cast<typename Allocator::Node*>(p);
                 uint8_t const expectedAge = pNode[-1].age;
-                FILAMENT_CHECK_POSTCONDITION(expectedAge == age) <<
-                        "use-after-free of Handle with id=" << handle.getId();
+                // getHandleTag() is only called if the check fails.
+                FILAMENT_CHECK_POSTCONDITION(expectedAge == age)
+                        << "use-after-free of Handle with id=" << handle.getId()
+                        << ", tag=" << getHandleTag(handle.getId()).c_str_safe();
+            }
+        } else {
+            // check for heap handle use-after-free
+            if (UTILS_UNLIKELY(!mUseAfterFreeCheckDisabled)) {
+                uint8_t const index = (handle.getId() & HANDLE_INDEX_MASK);
+                // if we've already handed out this handle index before, it's definitely a
+                // use-after-free, otherwise it's probably just a corrupted handle
+                if (index < mId) {
+                    FILAMENT_CHECK_POSTCONDITION(p != nullptr)
+                            << "use-after-free of heap Handle with id=" << handle.getId()
+                            << ", tag=" << getHandleTag(handle.getId()).c_str_safe();
+                } else {
+                    FILAMENT_CHECK_POSTCONDITION(p != nullptr)
+                            << "corrupted heap Handle with id=" << handle.getId()
+                            << ", tag=" << getHandleTag(handle.getId()).c_str_safe();
+                }
             }
         }
 
         return static_cast<Dp>(p);
     }
 
+    template<typename B>
+    bool is_valid(Handle<B>& handle) {
+        if (!handle) {
+            // null handles are invalid
+            return false;
+        }
+        auto [p, tag] = handleToPointer(handle.getId());
+        if (isPoolHandle(handle.getId())) {
+            uint8_t const age = (tag & HANDLE_AGE_MASK) >> HANDLE_AGE_SHIFT;
+            auto const pNode = static_cast<typename Allocator::Node*>(p);
+            uint8_t const expectedAge = pNode[-1].age;
+            return expectedAge == age;
+        }
+        return p != nullptr;
+    }
+
     template<typename Dp, typename B>
     inline typename std::enable_if_t<
             std::is_pointer_v<Dp> &&
@@ -189,6 +224,29 @@ public:
         return handle_cast<Dp>(const_cast<Handle<B>&>(handle));
     }
 
+    void associateTagToHandle(HandleBase::HandleId id, utils::CString&& tag) noexcept {
+        // TODO: for now, only pool handles check for use-after-free, so we only keep tags for
+        // those
+        if (isPoolHandle(id)) {
+            // Truncate the age to get the debug tag
+            uint32_t const key = id & ~(HANDLE_DEBUG_TAG_MASK ^ HANDLE_AGE_MASK);
+            // This line is the costly part. In the future, we could potentially use a custom
+            // allocator.
+            mDebugTags[key] = std::move(tag);
+        }
+    }
+
+    utils::CString getHandleTag(HandleBase::HandleId id) const noexcept {
+        if (!isPoolHandle(id)) {
+            return "(no tag)";
+        }
+        uint32_t const key = id & ~(HANDLE_DEBUG_TAG_MASK ^ HANDLE_AGE_MASK);
+        if (auto pos = mDebugTags.find(key); pos != mDebugTags.end()) {
+            return pos->second;
+        }
+        return "(no tag)";
+    }
+
 private:
 
     template<typename D>
@@ -306,12 +364,24 @@ private:
         }
     }
 
-    // we handle a 4 bits age per address
-    static constexpr uint32_t HANDLE_HEAP_FLAG      = 0x80000000u;      // pool vs heap handle
-    static constexpr uint32_t HANDLE_AGE_MASK       = 0x78000000u;      // handle's age
-    static constexpr uint32_t HANDLE_INDEX_MASK     = 0x07FFFFFFu;      // handle index
-    static constexpr uint32_t HANDLE_TAG_MASK       = HANDLE_AGE_MASK;
-    static constexpr uint32_t HANDLE_AGE_SHIFT      = 27;
+    // number if bits allotted to the handle's age (currently 4 max)
+    static constexpr uint32_t HANDLE_AGE_BIT_COUNT = 4;
+    // number if bits allotted to the handle's debug tag (HANDLE_AGE_BIT_COUNT max)
+    static constexpr uint32_t HANDLE_DEBUG_TAG_BIT_COUNT = 2;
+    // bit shift for both the age and debug tag
+    static constexpr uint32_t HANDLE_AGE_SHIFT = 27;
+    // mask for the heap (vs pool) flag
+    static constexpr uint32_t HANDLE_HEAP_FLAG = 0x80000000u;
+    // mask for the age
+    static constexpr uint32_t HANDLE_AGE_MASK =
+            ((1 << HANDLE_AGE_BIT_COUNT) - 1) << HANDLE_AGE_SHIFT;
+    // mask for the debug tag
+    static constexpr uint32_t HANDLE_DEBUG_TAG_MASK =
+            ((1 << HANDLE_DEBUG_TAG_BIT_COUNT) - 1) << HANDLE_AGE_SHIFT;
+    // mask for the index
+    static constexpr uint32_t HANDLE_INDEX_MASK = 0x07FFFFFFu;
+
+    static_assert(HANDLE_DEBUG_TAG_BIT_COUNT <= HANDLE_AGE_BIT_COUNT);
 
     static bool isPoolHandle(HandleBase::HandleId id) noexcept {
         return (id & HANDLE_HEAP_FLAG) == 0u;
@@ -326,7 +396,7 @@ private:
         // a non-pool handle.
         if (UTILS_LIKELY(isPoolHandle(id))) {
             char* const base = (char*)mHandleArena.getArea().begin();
-            uint32_t const tag = id & HANDLE_TAG_MASK;
+            uint32_t const tag = id & HANDLE_AGE_MASK;
             size_t const offset = (id & HANDLE_INDEX_MASK) * Allocator::getAlignment();
             return { static_cast<void*>(base + offset), tag };
         }
@@ -341,7 +411,7 @@ private:
         size_t const offset = (char*)p - base;
         assert_invariant((offset % Allocator::getAlignment()) == 0);
         auto id = HandleBase::HandleId(offset / Allocator::getAlignment());
-        id |= tag & HANDLE_TAG_MASK;
+        id |= tag & HANDLE_AGE_MASK;
         assert_invariant((id & HANDLE_HEAP_FLAG) == 0);
         return id;
     }
@@ -351,6 +421,7 @@ private:
     // Below is only used when running out of space in the HandleArena
     mutable utils::Mutex mLock;
     tsl::robin_map<HandleBase::HandleId, void*> mOverflowMap;
+    tsl::robin_map<HandleBase::HandleId, utils::CString> mDebugTags;
     HandleBase::HandleId mId = 0;
     bool mUseAfterFreeCheckDisabled = false;
 };

filament/backend/src/CircularBuffer.cpp

@@ -24,7 +24,7 @@
 #include <utils/debug.h>
 #include <utils/ostream.h>
 
-#if !defined(WIN32) && !defined(__EMSCRIPTEN__) && !defined(IOS)
+#if !defined(WIN32) && !defined(__EMSCRIPTEN__)
 #    include <sys/mman.h>
 #    include <unistd.h>
 #    define HAS_MMAP 1
@@ -32,10 +32,11 @@
 #    define HAS_MMAP 0
 #endif
 
-#include <stdint.h>
 #include <stddef.h>
-#include <stdlib.h>
+#include <stdint.h>
 #include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
 
 using namespace utils;
 
@@ -81,6 +82,9 @@ void* CircularBuffer::alloc(size_t size) noexcept {
             // map the circular buffer once...
             vaddr = mmap(reserve_vaddr, size, PROT_READ | PROT_WRITE, MAP_PRIVATE, fd, 0);
             if (vaddr != MAP_FAILED) {
+                // populate the address space with pages (because this is a circular buffer,
+                // all the pages will be allocated eventually, might as well do it now)
+                memset(vaddr, 0, size);
                 // and map the circular buffer again, behind the previous copy...
                 vaddr_shadow = mmap((char*)vaddr + size, size,
                         PROT_READ | PROT_WRITE, MAP_PRIVATE, fd, 0);
@@ -101,7 +105,7 @@ void* CircularBuffer::alloc(size_t size) noexcept {
     if (UTILS_UNLIKELY(mAshmemFd < 0)) {
         // ashmem failed
         if (vaddr_guard != MAP_FAILED) {
-            munmap(vaddr_guard, size);
+            munmap(vaddr_guard, BLOCK_SIZE);
         }
 
         if (vaddr_shadow != MAP_FAILED) {
@@ -119,12 +123,11 @@ void* CircularBuffer::alloc(size_t size) noexcept {
         data = mmap(nullptr, size * 2 + BLOCK_SIZE,
                 PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
 
-        FILAMENT_CHECK_POSTCONDITION(data) <<
+        FILAMENT_CHECK_POSTCONDITION(data != MAP_FAILED) <<
                 "couldn't allocate " << (size * 2 / 1024) <<
                 " KiB of virtual address space for the command buffer";
 
-        slog.d << "WARNING: Using soft CircularBuffer (" << (size * 2 / 1024) << " KiB)"
-               << io::endl;
+        slog.w << "Using 'soft' CircularBuffer (" << (size * 2 / 1024) << " KiB)" << io::endl;
 
         // guard page at the end
         void* guard = (void*)(uintptr_t(data) + size * 2);

filament/backend/src/CommandBufferQueue.cpp

@@ -101,9 +101,8 @@ void CommandBufferQueue::flush() noexcept {
     size_t const used = std::distance(
             static_cast<char const*>(begin), static_cast<char const*>(end));
 
+
     std::unique_lock<utils::Mutex> lock(mLock);
-    mCommandBuffersToExecute.push_back({ begin, end });
-    mCondition.notify_one();
 
     // circular buffer is too small, we corrupted the stream
     FILAMENT_CHECK_POSTCONDITION(used <= mFreeSpace) <<
@@ -112,10 +111,12 @@ void CommandBufferQueue::flush() noexcept {
             "Space used at this time: " << used <<
             " bytes, overflow: " << used - mFreeSpace << " bytes";
 
-    // wait until there is enough space in the buffer
     mFreeSpace -= used;
-    if (UTILS_UNLIKELY(mFreeSpace < requiredSize)) {
+    mCommandBuffersToExecute.push_back({ begin, end });
+    mCondition.notify_one();
 
+    // wait until there is enough space in the buffer
+    if (UTILS_UNLIKELY(mFreeSpace < requiredSize)) {
 
 #ifndef NDEBUG
         size_t const totalUsed = circularBuffer.size() - mFreeSpace;
@@ -153,8 +154,10 @@ std::vector<CommandBufferQueue::Range> CommandBufferQueue::waitForCommands() con
 }
 
 void CommandBufferQueue::releaseBuffer(CommandBufferQueue::Range const& buffer) {
+    size_t const used = std::distance(
+            static_cast<char const*>(buffer.begin), static_cast<char const*>(buffer.end));
     std::lock_guard<utils::Mutex> const lock(mLock);
-    mFreeSpace += uintptr_t(buffer.end) - uintptr_t(buffer.begin);
+    mFreeSpace += used;
     mCondition.notify_one();
 }
 

filament/backend/src/CommandStream.cpp

@@ -20,11 +20,16 @@
 #include <utils/CallStack.h>
 #endif
 
+#include <utils/compiler.h>
 #include <utils/Log.h>
+#include <utils/ostream.h>
 #include <utils/Profiler.h>
 #include <utils/Systrace.h>
 
+#include <cstddef>
 #include <functional>
+#include <string>
+#include <utility>
 
 #ifdef __ANDROID__
 #include <sys/system_properties.h>
@@ -74,8 +79,8 @@ CommandStream::CommandStream(Driver& driver, CircularBuffer& buffer) noexcept
 }
 
 void CommandStream::execute(void* buffer) {
-    SYSTRACE_CALL();
-    SYSTRACE_CONTEXT();
+    // NOTE: we can't use SYSTRACE_CALL() or similar here because, execute() below, also
+    // uses systrace BEGIN/END and the END is not guaranteed to be happening in this scope.
 
     Profiler profiler;
 
@@ -100,6 +105,7 @@ void CommandStream::execute(void* buffer) {
             // we want to remove all this when tracing is completely disabled
             profiler.stop();
             UTILS_UNUSED Profiler::Counters const counters = profiler.readCounters();
+            SYSTRACE_CONTEXT();
             SYSTRACE_VALUE32("GLThread (I)", counters.getInstructions());
             SYSTRACE_VALUE32("GLThread (C)", counters.getCpuCycles());
             SYSTRACE_VALUE32("GLThread (CPI x10)", counters.getCPI() * 10);

filament/backend/src/DriverBase.h

@@ -101,6 +101,14 @@ struct HwProgram : public HwBase {
     HwProgram() noexcept = default;
 };
 
+struct HwDescriptorSetLayout : public HwBase {
+    HwDescriptorSetLayout() noexcept = default;
+};
+
+struct HwDescriptorSet : public HwBase {
+    HwDescriptorSet() noexcept = default;
+};
+
 struct HwSamplerGroup : public HwBase {
     HwSamplerGroup() noexcept = default;
 };

filament/backend/src/HandleAllocator.cpp

@@ -80,6 +80,9 @@ HandleAllocator<P0, P1, P2>::HandleAllocator(const char* name, size_t size,
         bool disableUseAfterFreeCheck) noexcept
     : mHandleArena(name, size, disableUseAfterFreeCheck),
       mUseAfterFreeCheckDisabled(disableUseAfterFreeCheck) {
+    // Reserve initial space for debug tags. This prevents excessive calls to malloc when the first
+    // few tags are set.
+    mDebugTags.reserve(512);
 }
 
 template <size_t P0, size_t P1, size_t P2>

filament/backend/src/PlatformFactory.cpp

@@ -41,6 +41,10 @@
         #if defined(FILAMENT_SUPPORTS_OPENGL) && !defined(FILAMENT_USE_EXTERNAL_GLES3)
             #include "backend/platforms/PlatformEGLHeadless.h"
         #endif
+    #elif defined(FILAMENT_SUPPORTS_OSMESA)
+        #if defined(FILAMENT_SUPPORTS_OPENGL) && !defined(FILAMENT_USE_EXTERNAL_GLES3)
+            #include "backend/platforms/PlatformOSMesa.h"
+        #endif
     #endif
 #elif defined(WIN32)
     #if defined(FILAMENT_SUPPORTS_OPENGL) && !defined(FILAMENT_USE_EXTERNAL_GLES3)
@@ -124,6 +128,8 @@ Platform* PlatformFactory::create(Backend* backend) noexcept {
                 return new PlatformGLX();
             #elif defined(FILAMENT_SUPPORTS_EGL_ON_LINUX)
                 return new PlatformEGLHeadless();
+            #elif defined(FILAMENT_SUPPORTS_OSMESA)
+                return new PlatformOSMesa();
             #else
                 return nullptr;
             #endif

filament/backend/src/Program.cpp

@@ -14,7 +14,18 @@
  * limitations under the License.
  */
 
-#include "backend/Program.h"
+#include <backend/Program.h>
+#include <backend/DriverEnums.h>
+
+#include <utils/debug.h>
+#include <utils/CString.h>
+#include <utils/ostream.h>
+#include <utils/Invocable.h>
+
+#include <utility>
+
+#include <stddef.h>
+#include <stdint.h>
 
 namespace filament::backend {
 
@@ -52,41 +63,24 @@ Program& Program::shaderLanguage(ShaderLanguage shaderLanguage) {
     return *this;
 }
 
-Program& Program::uniformBlockBindings(
-        FixedCapacityVector<std::pair<utils::CString, uint8_t>> const& uniformBlockBindings) noexcept {
-    for (auto const& item : uniformBlockBindings) {
-        assert_invariant(item.second < UNIFORM_BINDING_COUNT);
-        mUniformBlocks[item.second] = item.first;
-    }
+Program& Program::descriptorBindings(backend::descriptor_set_t set,
+        DescriptorBindingsInfo descriptorBindings) noexcept {
+    mDescriptorBindings[set] = std::move(descriptorBindings);
     return *this;
 }
 
-Program& Program::uniforms(uint32_t index, UniformInfo const& uniforms) noexcept {
-    assert_invariant(index < UNIFORM_BINDING_COUNT);
-    mBindingUniformInfo[index] = uniforms;
+Program& Program::uniforms(uint32_t index, utils::CString name, UniformInfo uniforms) noexcept {
+    mBindingUniformsInfo.reserve(mBindingUniformsInfo.capacity() + 1);
+    mBindingUniformsInfo.emplace_back(index, std::move(name), std::move(uniforms));
     return *this;
 }
 
-
-Program& Program::attributes(
-        utils::FixedCapacityVector<std::pair<utils::CString, uint8_t>> attributes) noexcept {
+Program& Program::attributes(AttributesInfo attributes) noexcept {
     mAttributes = std::move(attributes);
     return *this;
 }
 
-Program& Program::setSamplerGroup(size_t bindingPoint, ShaderStageFlags stageFlags,
-        const Program::Sampler* samplers, size_t count) noexcept {
-    auto& groupData = mSamplerGroups[bindingPoint];
-    groupData.stageFlags = stageFlags;
-    auto& samplerList = groupData.samplers;
-    samplerList.reserve(count);
-    samplerList.resize(count);
-    std::copy_n(samplers, count, samplerList.data());
-    return *this;
-}
-
-Program& Program::specializationConstants(
-        FixedCapacityVector<SpecializationConstant> specConstants) noexcept {
+Program& Program::specializationConstants(SpecializationConstantsInfo specConstants) noexcept {
     mSpecializationConstants = std::move(specConstants);
     return *this;
 }

filament/backend/src/VirtualMachineEnv.cpp

@@ -16,13 +16,23 @@
 
 #include "private/backend/VirtualMachineEnv.h"
 
+#include <utils/compiler.h>
 #include <utils/debug.h>
 
+#include <jni.h>
+
 namespace filament {
 
 JavaVM* VirtualMachineEnv::sVirtualMachine = nullptr;
 
-// This is called when the library is loaded. We need this to get a reference to the global VM
+/*
+ * This is typically called by filament_jni.so when it is loaded. If filament_jni.so is not used,
+ * then this must be called manually -- however, this is a problem because VirtualMachineEnv.h
+ * is currently private and part of backend.
+ * For now, we authorize this usage, but we will need to fix it; by making a proper public
+ * API for this.
+ */
+UTILS_PUBLIC
 UTILS_NOINLINE
 jint VirtualMachineEnv::JNI_OnLoad(JavaVM* vm) noexcept {
     JNIEnv* env = nullptr;

filament/backend/src/metal/MetalBuffer.h

@@ -73,10 +73,11 @@ public:
     enum class Type {
         NONE = 0,
         GENERIC = 1,
-        RING = 2,
+        RING = 2,       // deprecated
         STAGING = 3,
+        DESCRIPTOR_SET = 4,
     };
-    static constexpr size_t TypeCount = 3;
+    static constexpr size_t TypeCount = 4;
 
     static constexpr auto toIndex(Type t) {
         assert_invariant(t != Type::NONE);
@@ -88,6 +89,8 @@ public:
                 return 1;
             case Type::STAGING:
                 return 2;
+            case Type::DESCRIPTOR_SET:
+                return 3;
         }
     }
 
@@ -160,6 +163,8 @@ public:
          size_t size, bool forceGpuBuffer = false);
     ~MetalBuffer();
 
+    [[nodiscard]] bool wasAllocationSuccessful() const noexcept { return mBuffer || mCpuBuffer; }
+
     MetalBuffer(const MetalBuffer& rhs) = delete;
     MetalBuffer& operator=(const MetalBuffer& rhs) = delete;
 
@@ -169,8 +174,10 @@ public:
      * Update the buffer with data inside src. Potentially allocates a new buffer allocation to hold
      * the bytes which will be released when the current frame is finished.
      */
-    void copyIntoBuffer(void* src, size_t size, size_t byteOffset);
-    void copyIntoBufferUnsynchronized(void* src, size_t size, size_t byteOffset);
+    using TagResolver = utils::Invocable<const char*(void)>;
+    void copyIntoBuffer(void* src, size_t size, size_t byteOffset, TagResolver&& getHandleTag);
+    void copyIntoBufferUnsynchronized(
+            void* src, size_t size, size_t byteOffset, TagResolver&& getHandleTag);
 
     /**
      * Denotes that this buffer is used for a draw call ensuring that its allocation remains valid
@@ -180,7 +187,7 @@ public:
      * is no device allocation.
      *
      */
-    id<MTLBuffer> getGpuBufferForDraw(id<MTLCommandBuffer> cmdBuffer) noexcept;
+    id<MTLBuffer> getGpuBufferForDraw() noexcept;
 
     void* getCpuBuffer() const noexcept { return mCpuBuffer; }
 
@@ -204,6 +211,17 @@ public:
 
 private:
 
+    enum class UploadStrategy {
+        POOL,
+        BUMP_ALLOCATOR,
+    };
+
+    void uploadWithPoolBuffer(
+            void* src, size_t size, size_t byteOffset, TagResolver&& getHandleTag) const;
+    void uploadWithBumpAllocator(
+            void* src, size_t size, size_t byteOffset, TagResolver&& getHandleTag) const;
+
+    UploadStrategy mUploadStrategy;
     TrackedMetalBuffer mBuffer;
     size_t mBufferSize = 0;
     void* mCpuBuffer = nullptr;

filament/backend/src/metal/MetalBuffer.mm

@@ -27,16 +27,28 @@ MetalPlatform* TrackedMetalBuffer::platform = nullptr;
 MetalPlatform* ScopedAllocationTimer::platform = nullptr;
 
 MetalBuffer::MetalBuffer(MetalContext& context, BufferObjectBinding bindingType, BufferUsage usage,
-        size_t size, bool forceGpuBuffer) : mBufferSize(size), mContext(context) {
+        size_t size, bool forceGpuBuffer)
+    : mBufferSize(size), mContext(context) {
+    const MetalBumpAllocator& allocator = *mContext.bumpAllocator;
+    // VERTEX is also used for index buffers
+    if (allocator.getCapacity() > 0 && bindingType == BufferObjectBinding::VERTEX) {
+        mUploadStrategy = UploadStrategy::BUMP_ALLOCATOR;
+    } else {
+        mUploadStrategy = UploadStrategy::POOL;
+    }
+
     // If the buffer is less than 4K in size and is updated frequently, we don't use an explicit
     // buffer. Instead, we use immediate command encoder methods like setVertexBytes:length:atIndex:.
     // This won't work for SSBOs, since they are read/write.
+
+    /*
     if (size <= 4 * 1024 && bindingType != BufferObjectBinding::SHADER_STORAGE &&
             usage == BufferUsage::DYNAMIC && !forceGpuBuffer) {
         mBuffer = nil;
         mCpuBuffer = malloc(size);
         return;
     }
+    */
 
     // Otherwise, we allocate a private GPU buffer.
     {
@@ -44,8 +56,8 @@ MetalBuffer::MetalBuffer(MetalContext& context, BufferObjectBinding bindingType,
         mBuffer = { [context.device newBufferWithLength:size options:MTLResourceStorageModePrivate],
             TrackedMetalBuffer::Type::GENERIC };
     }
-    FILAMENT_CHECK_POSTCONDITION(mBuffer)
-            << "Could not allocate Metal buffer of size " << size << ".";
+    // mBuffer might fail to be allocated. Clients can check for this by calling
+    // wasAllocationSuccessful().
 }
 
 MetalBuffer::~MetalBuffer() {
@@ -54,49 +66,44 @@ MetalBuffer::~MetalBuffer() {
     }
 }
 
-void MetalBuffer::copyIntoBuffer(void* src, size_t size, size_t byteOffset) {
+void MetalBuffer::copyIntoBuffer(
+        void* src, size_t size, size_t byteOffset, TagResolver&& getHandleTag) {
     if (size <= 0) {
         return;
     }
+
+    FILAMENT_CHECK_PRECONDITION(src)
+            << "copyIntoBuffer called with a null src, tag=" << getHandleTag();
     FILAMENT_CHECK_PRECONDITION(size + byteOffset <= mBufferSize)
             << "Attempting to copy " << size << " bytes into a buffer of size " << mBufferSize
-            << " at offset " << byteOffset;
+            << " at offset " << byteOffset << ", tag=" << getHandleTag();
+    // The copy blit requires that byteOffset be a multiple of 4.
+    FILAMENT_CHECK_PRECONDITION(!(byteOffset & 0x3))
+            << "byteOffset must be a multiple of 4, tag=" << getHandleTag();
 
-    // Either copy into the Metal buffer or into our cpu buffer.
+    // If we have a cpu buffer, we can directly copy into it.
     if (mCpuBuffer) {
         memcpy(static_cast<uint8_t*>(mCpuBuffer) + byteOffset, src, size);
         return;
     }
 
-    // Acquire a staging buffer to hold the contents of this update.
-    MetalBufferPool* bufferPool = mContext.bufferPool;
-    const MetalBufferPoolEntry* const staging = bufferPool->acquireBuffer(size);
-    memcpy(staging->buffer.get().contents, src, size);
-
-    // The blit below requires that byteOffset be a multiple of 4.
-    FILAMENT_CHECK_PRECONDITION(!(byteOffset & 0x3u)) << "byteOffset must be a multiple of 4";
-
-    // Encode a blit from the staging buffer into the private GPU buffer.
-    id<MTLCommandBuffer> cmdBuffer = getPendingCommandBuffer(&mContext);
-    id<MTLBlitCommandEncoder> blitEncoder = [cmdBuffer blitCommandEncoder];
-    blitEncoder.label = @"Buffer upload blit";
-    [blitEncoder copyFromBuffer:staging->buffer.get()
-                   sourceOffset:0
-                       toBuffer:mBuffer.get()
-              destinationOffset:byteOffset
-                           size:size];
-    [blitEncoder endEncoding];
-    [cmdBuffer addCompletedHandler:^(id<MTLCommandBuffer> cb) {
-        bufferPool->releaseBuffer(staging);
-    }];
+    switch (mUploadStrategy) {
+        case UploadStrategy::BUMP_ALLOCATOR:
+            uploadWithBumpAllocator(src, size, byteOffset, std::move(getHandleTag));
+            break;
+        case UploadStrategy::POOL:
+            uploadWithPoolBuffer(src, size, byteOffset, std::move(getHandleTag));
+            break;
+    }
 }
 
-void MetalBuffer::copyIntoBufferUnsynchronized(void* src, size_t size, size_t byteOffset) {
+void MetalBuffer::copyIntoBufferUnsynchronized(
+        void* src, size_t size, size_t byteOffset, TagResolver&& getHandleTag) {
     // TODO: implement the unsynchronized version
-    copyIntoBuffer(src, size, byteOffset);
+    copyIntoBuffer(src, size, byteOffset, std::move(getHandleTag));
 }
 
-id<MTLBuffer> MetalBuffer::getGpuBufferForDraw(id<MTLCommandBuffer> cmdBuffer) noexcept {
+id<MTLBuffer> MetalBuffer::getGpuBufferForDraw() noexcept {
     // If there's a CPU buffer, then we return nil here, as the CPU-side buffer will be bound
     // separately.
     if (mCpuBuffer) {
@@ -139,7 +146,7 @@ void MetalBuffer::bindBuffers(id<MTLCommandBuffer> cmdBuffer, id<MTLCommandEncod
         }
         // getGpuBufferForDraw() might return nil, which means there isn't a device allocation for
         // this buffer. In this case, we'll bind the buffer below with the CPU-side memory.
-        id<MTLBuffer> gpuBuffer = buffer->getGpuBufferForDraw(cmdBuffer);
+        id<MTLBuffer> gpuBuffer = buffer->getGpuBufferForDraw();
         if (!gpuBuffer) {
             continue;
         }
@@ -199,5 +206,54 @@ void MetalBuffer::bindBuffers(id<MTLCommandBuffer> cmdBuffer, id<MTLCommandEncod
     }
 }
 
+void MetalBuffer::uploadWithPoolBuffer(
+        void* src, size_t size, size_t byteOffset, TagResolver&& getHandleTag) const {
+    MetalBufferPool* bufferPool = mContext.bufferPool;
+    const MetalBufferPoolEntry* const staging = bufferPool->acquireBuffer(size);
+    FILAMENT_CHECK_POSTCONDITION(staging)
+            << "uploadWithPoolbuffer unable to acquire staging buffer of size " << size
+            << ", tag=" << getHandleTag();
+    memcpy(staging->buffer.get().contents, src, size);
+
+    // Encode a blit from the staging buffer into the private GPU buffer.
+    id<MTLCommandBuffer> cmdBuffer = getPendingCommandBuffer(&mContext);
+    id<MTLBlitCommandEncoder> blitEncoder = [cmdBuffer blitCommandEncoder];
+    blitEncoder.label = @"Buffer upload blit - pool buffer";
+    [blitEncoder copyFromBuffer:staging->buffer.get()
+                   sourceOffset:0
+                       toBuffer:mBuffer.get()
+              destinationOffset:byteOffset
+                           size:size];
+    [blitEncoder endEncoding];
+    [cmdBuffer addCompletedHandler:^(id<MTLCommandBuffer> cb) {
+        bufferPool->releaseBuffer(staging);
+    }];
+}
+
+void MetalBuffer::uploadWithBumpAllocator(
+        void* src, size_t size, size_t byteOffset, TagResolver&& getHandleTag) const {
+    MetalBumpAllocator& allocator = *mContext.bumpAllocator;
+    auto [buffer, offset] = allocator.allocateStagingArea(size);
+    FILAMENT_CHECK_POSTCONDITION(buffer)
+            << "uploadWithBumpAllocator unable to acquire staging area of size " << size
+            << ", tag=" << getHandleTag();
+    void* const contents = buffer.contents;
+    FILAMENT_CHECK_POSTCONDITION(contents)
+            << "uploadWithBumpAllocator unable to acquire pointer to staging area, size " << size
+            << ", tag=" << getHandleTag();
+    memcpy(static_cast<char*>(contents) + offset, src, size);
+
+    // Encode a blit from the staging buffer into the private GPU buffer.
+    id<MTLCommandBuffer> cmdBuffer = getPendingCommandBuffer(&mContext);
+    id<MTLBlitCommandEncoder> blitEncoder = [cmdBuffer blitCommandEncoder];
+    blitEncoder.label = @"Buffer upload blit - bump allocator";
+    [blitEncoder copyFromBuffer:buffer
+                   sourceOffset:offset
+                       toBuffer:mBuffer.get()
+              destinationOffset:byteOffset
+                           size:size];
+    [blitEncoder endEncoding];
+}
+
 } // namespace backend
 } // namespace filament

filament/backend/src/metal/MetalBufferPool.h

@@ -38,6 +38,28 @@ struct MetalBufferPoolEntry {
     mutable uint32_t referenceCount;
 };
 
+class MetalBumpAllocator {
+public:
+    MetalBumpAllocator(id<MTLDevice> device, size_t capacity);
+
+    /**
+     * Allocates a staging area of the given size. Returns a pair of the buffer and the offset
+     * within the buffer. The buffer is guaranteed to be at least the given size, but may be larger.
+     * Clients must not write to the buffer beyond the returned offset + size.
+     * Clients are responsible for holding a reference to the returned buffer.
+     * Allocations are guaranteed to be aligned to 4 bytes.
+     */
+    std::pair<id<MTLBuffer>, size_t> allocateStagingArea(size_t size);
+
+    size_t getCapacity() const noexcept { return mCapacity; }
+
+private:
+    id<MTLDevice> mDevice;
+    TrackedMetalBuffer mCurrentUploadBuffer = nil;
+    size_t mHead = 0;
+    size_t mCapacity;
+};
+
 // Manages a pool of Metal buffers, periodically releasing ones that have been unused for awhile.
 class MetalBufferPool {
 public:

filament/backend/src/metal/MetalBufferPool.mm

@@ -116,5 +116,39 @@ void MetalBufferPool::reset() noexcept {
     mFreeStages.clear();
 }
 
+MetalBumpAllocator::MetalBumpAllocator(id<MTLDevice> device, size_t capacity)
+    : mDevice(device), mCapacity(capacity) {
+    if (mCapacity > 0) {
+        mCurrentUploadBuffer = { [device newBufferWithLength:capacity options:MTLStorageModeShared],
+            TrackedMetalBuffer::Type::STAGING };
+    }
+}
+
+std::pair<id<MTLBuffer>, size_t> MetalBumpAllocator::allocateStagingArea(size_t size) {
+    if (size == 0) {
+        return { nil, 0 };
+    }
+    if (size > mCapacity) {
+        return { [mDevice newBufferWithLength:size options:MTLStorageModeShared], 0 };
+    }
+    assert_invariant(mCurrentUploadBuffer);
+
+    // Align the head to a 4-byte boundary.
+    mHead = (mHead + 3) & ~3;
+
+    if (UTILS_LIKELY(mHead + size <= mCapacity)) {
+        const size_t oldHead = mHead;
+        mHead += size;
+        return { mCurrentUploadBuffer.get(), oldHead };
+    }
+
+    // We're finished with the current allocation.
+    mCurrentUploadBuffer = { [mDevice newBufferWithLength:mCapacity options:MTLStorageModeShared],
+        TrackedMetalBuffer::Type::STAGING };
+    mHead = size;
+
+    return { mCurrentUploadBuffer.get(), 0 };
+}
+
 } // namespace backend
 } // namespace filament

filament/backend/src/metal/MetalContext.h

@@ -21,6 +21,8 @@
 #include "MetalShaderCompiler.h"
 #include "MetalState.h"
 
+#include <backend/DriverEnums.h>
+
 #include <CoreVideo/CVMetalTextureCache.h>
 #include <Metal/Metal.h>
 #include <QuartzCore/QuartzCore.h>
@@ -44,14 +46,15 @@ namespace backend {
 class MetalDriver;
 class MetalBlitter;
 class MetalBufferPool;
+class MetalBumpAllocator;
 class MetalRenderTarget;
-class MetalSamplerGroup;
 class MetalSwapChain;
 class MetalTexture;
 class MetalTimerQueryInterface;
 struct MetalUniformBuffer;
 struct MetalIndexBuffer;
 struct MetalVertexBuffer;
+struct MetalDescriptorSet;
 
 constexpr static uint8_t MAX_SAMPLE_COUNT = 8;  // Metal devices support at most 8 MSAA samples
 
@@ -67,16 +70,64 @@ private:
     bool mDirty = false;
 };
 
-struct MetalContext {
-    explicit MetalContext(size_t metalFreedTextureListSize)
-        : texturesToDestroy(metalFreedTextureListSize) {}
+class MetalDynamicOffsets {
+public:
+    void setOffsets(uint32_t set, const uint32_t* offsets, uint32_t count) {
+        assert(set < MAX_DESCRIPTOR_SET_COUNT);
 
+        auto getStartIndexForSet = [&](uint32_t s) {
+            uint32_t startIndex = 0;
+            for (uint32_t i = 0; i < s; i++) {
+                startIndex += mCounts[i];
+            }
+            return startIndex;
+        };
+
+        const bool resizeNecessary = mCounts[set] != count;
+        if (UTILS_UNLIKELY(resizeNecessary)) {
+            int delta = count - mCounts[set];
+
+            auto thisSetStart = mOffsets.begin() + getStartIndexForSet(set);
+            if (delta > 0) {
+                mOffsets.insert(thisSetStart, delta, 0);
+            } else {
+                mOffsets.erase(thisSetStart, thisSetStart - delta);
+            }
+
+            mCounts[set] = count;
+        }
+
+        if (resizeNecessary ||
+                !std::equal(
+                        offsets, offsets + count, mOffsets.begin() + getStartIndexForSet(set))) {
+            std::copy(offsets, offsets + count, mOffsets.begin() + getStartIndexForSet(set));
+            mDirty = true;
+        }
+    }
+    bool isDirty() const { return mDirty; }
+    void setDirty(bool dirty) { mDirty = dirty; }
+
+    std::pair<uint32_t, const uint32_t*> getOffsets() const {
+        return { mOffsets.size(), mOffsets.data() };
+    }
+
+private:
+    std::array<uint32_t, MAX_DESCRIPTOR_SET_COUNT> mCounts = { 0 };
+    std::vector<uint32_t> mOffsets;
+    bool mDirty = false;
+};
+
+struct MetalContext {
     MetalDriver* driver;
     id<MTLDevice> device = nullptr;
     id<MTLCommandQueue> commandQueue = nullptr;
 
-    id<MTLCommandBuffer> pendingCommandBuffer = nullptr;
-    id<MTLRenderCommandEncoder> currentRenderPassEncoder = nullptr;
+    // The ID of pendingCommandBuffer (or the next command buffer, if pendingCommandBuffer is nil).
+    uint64_t pendingCommandBufferId = 1;
+    // read from driver thread, set from completion handlers
+    std::atomic<uint64_t> latestCompletedCommandBufferId = 0;
+    id<MTLCommandBuffer> pendingCommandBuffer = nil;
+    id<MTLRenderCommandEncoder> currentRenderPassEncoder = nil;
 
     std::atomic<bool> memorylessLimitsReached = false;
 
@@ -92,7 +143,7 @@ struct MetalContext {
     } highestSupportedGpuFamily;
 
     struct {
-        bool a8xStaticTextureTargetError;
+        bool staticTextureTargetError;
     } bugs;
 
     // sampleCountLookup[requestedSamples] gives a <= sample count supported by the device.
@@ -107,10 +158,9 @@ struct MetalContext {
     // State trackers.
     PipelineStateTracker pipelineState;
     DepthStencilStateTracker depthStencilState;
-    std::array<BufferState, Program::UNIFORM_BINDING_COUNT> uniformState;
-    std::array<BufferState, MAX_SSBO_COUNT> ssboState;
     CullModeStateTracker cullModeState;
     WindingStateTracker windingState;
+    DepthClampStateTracker depthClampState;
     Handle<HwRenderPrimitive> currentRenderPrimitive;
 
     // State caches.
@@ -123,24 +173,19 @@ struct MetalContext {
 
     std::array<MetalPushConstantBuffer, Program::SHADER_TYPE_COUNT> currentPushConstants;
 
-    MetalSamplerGroup* samplerBindings[Program::SAMPLER_BINDING_COUNT] = {};
+    // Keeps track of descriptor sets we've finalized for the current render pass.
+    tsl::robin_set<MetalDescriptorSet*> finalizedDescriptorSets;
+    std::array<MetalDescriptorSet*, MAX_DESCRIPTOR_SET_COUNT> currentDescriptorSets = {};
+    MetalBufferBindings<MAX_DESCRIPTOR_SET_COUNT, ShaderStage::VERTEX> vertexDescriptorBindings;
+    MetalBufferBindings<MAX_DESCRIPTOR_SET_COUNT, ShaderStage::FRAGMENT> fragmentDescriptorBindings;
+    MetalBufferBindings<MAX_DESCRIPTOR_SET_COUNT, ShaderStage::COMPUTE> computeDescriptorBindings;
+    MetalDynamicOffsets dynamicOffsets;
 
-    // Keeps track of sampler groups we've finalized for the current render pass.
-    tsl::robin_set<MetalSamplerGroup*> finalizedSamplerGroups;
-
-    // Keeps track of all alive sampler groups, textures.
-    tsl::robin_set<MetalSamplerGroup*> samplerGroups;
+    // Keeps track of all alive textures.
     tsl::robin_set<MetalTexture*> textures;
 
-    // This circular buffer implements delayed destruction for Metal texture handles. It keeps a
-    // handle to a fixed number of the most recently destroyed texture handles. When we're asked to
-    // destroy a texture handle, we free its texture memory, but keep the MetalTexture object alive,
-    // marking it as "terminated". If we later are asked to use that texture, we can check its
-    // terminated status and throw an Objective-C error instead of crashing, which is helpful for
-    // debugging use-after-free issues in release builds.
-    utils::FixedCircularBuffer<Handle<HwTexture>> texturesToDestroy;
-
     MetalBufferPool* bufferPool;
+    MetalBumpAllocator* bumpAllocator;
 
     MetalSwapChain* currentDrawSwapChain = nil;
     MetalSwapChain* currentReadSwapChain = nil;
@@ -151,6 +196,7 @@ struct MetalContext {
 
     // Empty texture used to prevent GPU errors when a sampler has been bound without a texture.
     id<MTLTexture> emptyTexture = nil;
+    id<MTLBuffer> emptyBuffer = nil;
 
     MetalBlitter* blitter = nullptr;
 

filament/backend/src/metal/MetalContext.mm

@@ -101,9 +101,14 @@ id<MTLCommandBuffer> getPendingCommandBuffer(MetalContext* context) {
     context->pendingCommandBuffer = [context->commandQueue commandBuffer];
     // It's safe for this block to capture the context variable. MetalDriver::terminate will ensure
     // all frames and their completion handlers finish before context is deallocated.
+    uint64_t thisCommandBufferId = context->pendingCommandBufferId;
     [context->pendingCommandBuffer addCompletedHandler:^(id <MTLCommandBuffer> buffer) {
         context->resourceTracker.clearResources((__bridge void*) buffer);
-        
+
+        // Command buffers should complete in order, so latestCompletedCommandBufferId will only
+        // ever increase.
+        context->latestCompletedCommandBufferId = thisCommandBufferId;
+
         auto errorCode = (MTLCommandBufferError)buffer.error.code;
         if (@available(macOS 11.0, *)) {
             if (errorCode == MTLCommandBufferErrorMemoryless) {
@@ -125,6 +130,7 @@ void submitPendingCommands(MetalContext* context) {
     assert_invariant(context->pendingCommandBuffer.status != MTLCommandBufferStatusCommitted);
     [context->pendingCommandBuffer commit];
     context->pendingCommandBuffer = nil;
+    context->pendingCommandBufferId++;
 }
 
 id<MTLTexture> getOrCreateEmptyTexture(MetalContext* context) {
@@ -167,7 +173,6 @@ void MetalPushConstantBuffer::setPushConstant(PushConstantVariant value, uint8_t
 
 void MetalPushConstantBuffer::setBytes(id<MTLCommandEncoder> encoder, ShaderStage stage) {
     constexpr size_t PUSH_CONSTANT_SIZE_BYTES = 4;
-    constexpr size_t PUSH_CONSTANT_BUFFER_INDEX = 26;
 
     static char buffer[MAX_PUSH_CONSTANT_COUNT * PUSH_CONSTANT_SIZE_BYTES];
     assert_invariant(mPushConstants.size() <= MAX_PUSH_CONSTANT_COUNT);

filament/backend/src/metal/MetalDriver.h

@@ -32,6 +32,7 @@
 #include <functional>
 #include <mutex>
 #include <vector>
+#include <deque>
 
 namespace filament {
 namespace backend {
@@ -57,11 +58,11 @@ class MetalDriver final : public DriverBase {
 
 public:
     static Driver* create(MetalPlatform* platform, const Platform::DriverConfig& driverConfig);
-    void runAtNextTick(const std::function<void()>& fn) noexcept;
 
 private:
 
     friend class MetalSwapChain;
+    friend struct MetalDescriptorSet;
 
     MetalPlatform& mPlatform;
     MetalContext* mContext;
@@ -73,10 +74,23 @@ private:
 
     /*
      * Tasks run regularly on the driver thread.
+     * Not thread-safe; tasks are run from the driver thead and must be enqueued from the driver
+     * thread.
      */
+    void runAtNextTick(const std::function<void()>& fn) noexcept;
     void executeTickOps() noexcept;
     std::vector<std::function<void()>> mTickOps;
-    std::mutex mTickOpsLock;
+
+    // Tasks regularly executed on the driver thread after a command buffer has completed
+    struct DeferredTask {
+        DeferredTask(uint64_t commandBufferId, utils::Invocable<void()>&& fn) noexcept
+            : commandBufferId(commandBufferId), fn(std::move(fn)) {}
+        uint64_t commandBufferId;     // after this command buffer completes
+        utils::Invocable<void()> fn;  // execute this task
+    };
+    void executeAfterCurrentCommandBufferCompletes(utils::Invocable<void()>&& fn) noexcept;
+    void executeDeferredOps() noexcept;
+    std::deque<DeferredTask> mDeferredTasks;
 
     /*
      * Driver interface
@@ -137,7 +151,6 @@ private:
     inline void setRenderPrimitiveBuffer(Handle<HwRenderPrimitive> rph, PrimitiveType pt,
             Handle<HwVertexBuffer> vbh, Handle<HwIndexBuffer> ibh);
 
-    void finalizeSamplerGroup(MetalSamplerGroup* sg);
     void enumerateBoundBuffers(BufferObjectBinding bindingType,
             const std::function<void(const BufferState&, MetalBuffer*, uint32_t)>& f);
 

filament/backend/src/metal/MetalExternalImage.h

@@ -32,100 +32,75 @@ struct MetalContext;
  * texture.
  */
 class MetalExternalImage {
-
 public:
+    MetalExternalImage() = default;
 
-    MetalExternalImage(MetalContext& context,
-            TextureSwizzle r = TextureSwizzle::CHANNEL_0,
-            TextureSwizzle g = TextureSwizzle::CHANNEL_1,
-            TextureSwizzle b = TextureSwizzle::CHANNEL_2,
-            TextureSwizzle a = TextureSwizzle::CHANNEL_3) noexcept;
+    MetalExternalImage(MetalExternalImage&&);
+    MetalExternalImage& operator=(MetalExternalImage&&);
+    ~MetalExternalImage() noexcept;
+
+    MetalExternalImage(const MetalExternalImage&) = delete;
+    MetalExternalImage& operator=(const MetalExternalImage&) = delete;
 
     /**
-     * @return true, if this MetalExternalImage is holding a live external image. Returns false
-     * until set has been called with a valid CVPixelBuffer. The image can be cleared via
-     * set(nullptr), and isValid will return false again.
+     * While the texture is used for rendering, this MetalExternalImage must be kept alive.
      */
-    bool isValid() const noexcept;
+    id<MTLTexture> getMtlTexture() const noexcept;
+
+    bool isValid() const noexcept {
+        return mImage != nil || mRgbTexture != nullptr;
+    }
+
+    NSUInteger getWidth() const noexcept;
+    NSUInteger getHeight() const noexcept;
 
     /**
-     * Set this external image to the passed-in CVPixelBuffer. Future calls to
-     * getMetalTextureForDraw will return a texture backed by this CVPixelBuffer. Previous
-     * CVPixelBuffers and related resources will be released when all GPU work using them has
-     * finished.
+     * Create an external image with the passed-in CVPixelBuffer.
      *
-     * Calling set with a YCbCr image will encode a compute pass to convert the image from YCbCr to
-     * RGB.
+     * Ownership is taken of the CVPixelBuffer, which will be released when the returned
+     * MetalExternalImage is destroyed (or, in the case of a YCbCr image, after the conversion has
+     * completed).
+     *
+     * Calling set with a YCbCr image will encode a compute pass to convert the image from
+     * YCbCr to RGB.
      */
-    void set(CVPixelBufferRef image) noexcept;
+    static MetalExternalImage createFromImage(MetalContext& context, CVPixelBufferRef image);
 
     /**
-     * Set this external image to a specific plane of the passed-in CVPixelBuffer. Future calls to
-     * getMetalTextureForDraw will return a texture backed by a single plane of this CVPixelBuffer.
-     * Previous CVPixelBuffers and related resources will be released when all GPU work using them
-     * has finished.
+     * Create an external image with a specific plane of the passed-in CVPixelBuffer.
+     *
+     * Ownership is taken of the CVPixelBuffer, which will be released when the returned
+     * MetalExternalImage is destroyed.
      */
-    void set(CVPixelBufferRef image, size_t plane) noexcept;
+    static MetalExternalImage createFromImagePlane(
+            MetalContext& context, CVPixelBufferRef image, uint32_t plane);
 
-    /**
-     * Returns the width of the external image, or 0 if one is not set. For YCbCr images, returns
-     * the width of the luminance plane.
-     */
-    size_t getWidth() const noexcept { return mWidth; }
-
-    /**
-     * Returns the height of the external image, or 0 if one is not set. For YCbCr images, returns
-     * the height of the luminance plane.
-     */
-    size_t getHeight() const noexcept { return mHeight; }
-
-    /**
-     * Get a Metal texture used to draw this image and denote that it is used for the current frame.
-     * For future frames that use this external image, getMetalTextureForDraw must be called again.
-     */
-    id<MTLTexture> getMetalTextureForDraw() const noexcept;
+    static void assertWritableImage(CVPixelBufferRef image);
 
     /**
      * Free resources. Should be called at least once when no further calls to set will occur.
      */
     static void shutdown(MetalContext& context) noexcept;
 
-    static void assertWritableImage(CVPixelBufferRef image);
-
 private:
+    MetalExternalImage(CVPixelBufferRef image, CVMetalTextureRef texture) noexcept
+        : mImage(image), mTexture(texture) {}
+    explicit MetalExternalImage(id<MTLTexture> texture) noexcept : mRgbTexture(texture) {}
 
-    void unset();
-
-    CVMetalTextureRef createTextureFromImage(CVPixelBufferRef image, MTLPixelFormat format,
-            size_t plane);
-    id<MTLTexture> createRgbTexture(size_t width, size_t height);
-    id<MTLTexture> createSwizzledTextureView(id<MTLTexture> texture) const;
-    id<MTLTexture> createSwizzledTextureView(CVMetalTextureRef texture) const;
-    void ensureComputePipelineState();
-    id<MTLCommandBuffer> encodeColorConversionPass(id<MTLTexture> inYPlane, id<MTLTexture>
-            inCbCrTexture, id<MTLTexture> outTexture);
+    static id<MTLTexture> createRgbTexture(id<MTLDevice> device, size_t width, size_t height);
+    static CVMetalTextureRef createTextureFromImage(CVMetalTextureCacheRef textureCache,
+            CVPixelBufferRef image, MTLPixelFormat format, size_t plane);
+    static void ensureComputePipelineState(MetalContext& context);
+    static id<MTLCommandBuffer> encodeColorConversionPass(MetalContext& context,
+            id<MTLTexture> inYPlane, id<MTLTexture> inCbCrTexture, id<MTLTexture> outTexture);
 
     static constexpr size_t Y_PLANE = 0;
     static constexpr size_t CBCR_PLANE = 1;
 
-    MetalContext& mContext;
-
-    // If the external image has a single plane, mImage and mTexture hold references to the image
-    // and created Metal texture, respectively.
-    // mTextureView is a view of mTexture with any swizzling applied.
+    // TODO: this could probably be a union.
     CVPixelBufferRef mImage = nullptr;
     CVMetalTextureRef mTexture = nullptr;
-    id<MTLTexture> mTextureView = nullptr;
-    size_t mWidth = 0;
-    size_t mHeight = 0;
-
-    // If the external image is in the YCbCr format, this holds the result of the converted RGB
-    // texture.
     id<MTLTexture> mRgbTexture = nil;
-
-    struct {
-        TextureSwizzle r, g, b, a;
-    } mSwizzle;
 };
 
 } // namespace backend

filament/backend/src/metal/MetalExternalImage.mm

@@ -34,10 +34,6 @@
 namespace filament {
 namespace backend {
 
-static const auto cvBufferDeleter = [](const void* buffer) {
-    CVBufferRelease((CVMetalTextureRef) buffer);
-};
-
 static const char* kernel = R"(
 #include <metal_stdlib>
 #include <simd/simd.h>
@@ -71,18 +67,30 @@ ycbcrToRgb(texture2d<half, access::read>  inYTexture    [[texture(0)]],
 }
 )";
 
-MetalExternalImage::MetalExternalImage(MetalContext& context, TextureSwizzle r, TextureSwizzle g,
-        TextureSwizzle b, TextureSwizzle a) noexcept : mContext(context), mSwizzle{r, g, b, a} { }
-
-bool MetalExternalImage::isValid() const noexcept {
-    return mRgbTexture != nil || mImage != nullptr;
+NSUInteger MetalExternalImage::getWidth() const noexcept {
+    if (mImage) {
+        return CVPixelBufferGetWidth(mImage);
+    }
+    if (mRgbTexture) {
+        return mRgbTexture.width;
+    }
+    return 0;
 }
 
-void MetalExternalImage::set(CVPixelBufferRef image) noexcept {
-    unset();
+NSUInteger MetalExternalImage::getHeight() const noexcept {
+    if (mImage) {
+        return CVPixelBufferGetHeight(mImage);
+    }
+    if (mRgbTexture) {
+        return mRgbTexture.height;
+    }
+    return 0;
+}
 
+MetalExternalImage MetalExternalImage::createFromImage(
+        MetalContext& context, CVPixelBufferRef image) {
     if (!image) {
-        return;
+        return {};
     }
 
     OSType formatType = CVPixelBufferGetPixelFormatType(image);
@@ -96,30 +104,29 @@ void MetalExternalImage::set(CVPixelBufferRef image) noexcept {
             << ".";
 
     if (planeCount == 0) {
-        mImage = image;
-        mTexture = createTextureFromImage(image, MTLPixelFormatBGRA8Unorm, 0);
-        mTextureView = createSwizzledTextureView(mTexture);
-        mWidth = CVPixelBufferGetWidth(image);
-        mHeight = CVPixelBufferGetHeight(image);
+        CVMetalTextureRef texture =
+                createTextureFromImage(context.textureCache, image, MTLPixelFormatBGRA8Unorm, 0);
+        return { CVPixelBufferRetain(image), texture };
     }
 
     if (planeCount == 2) {
-        CVMetalTextureRef yPlane = createTextureFromImage(image, MTLPixelFormatR8Unorm, Y_PLANE);
-        CVMetalTextureRef cbcrPlane = createTextureFromImage(image, MTLPixelFormatRG8Unorm,
-                CBCR_PLANE);
+        CVPixelBufferRetain(image);
+
+        CVMetalTextureRef yPlane =
+                createTextureFromImage(context.textureCache, image, MTLPixelFormatR8Unorm, Y_PLANE);
+        CVMetalTextureRef cbcrPlane =
+                createTextureFromImage(context.textureCache, image, MTLPixelFormatRG8Unorm, CBCR_PLANE);
 
         // Get the size of luminance plane.
-        mWidth = CVPixelBufferGetWidthOfPlane(image, Y_PLANE);
-        mHeight = CVPixelBufferGetHeightOfPlane(image, Y_PLANE);
+        NSUInteger width = CVPixelBufferGetWidthOfPlane(image, Y_PLANE);
+        NSUInteger height = CVPixelBufferGetHeightOfPlane(image, Y_PLANE);
 
-        id<MTLTexture> rgbTexture = createRgbTexture(mWidth, mHeight);
-        id<MTLCommandBuffer> commandBuffer = encodeColorConversionPass(
+        id<MTLTexture> rgbTexture = createRgbTexture(context.device, width, height);
+        id<MTLCommandBuffer> commandBuffer = encodeColorConversionPass(context,
                 CVMetalTextureGetTexture(yPlane),
                 CVMetalTextureGetTexture(cbcrPlane),
                 rgbTexture);
 
-        mRgbTexture = createSwizzledTextureView(rgbTexture);
-
         [commandBuffer addCompletedHandler:^(id <MTLCommandBuffer> o) {
             CVBufferRelease(yPlane);
             CVBufferRelease(cbcrPlane);
@@ -127,70 +134,83 @@ void MetalExternalImage::set(CVPixelBufferRef image) noexcept {
         }];
 
         [commandBuffer commit];
+        return MetalExternalImage { rgbTexture };
     }
+
+    return {};
 }
 
-void MetalExternalImage::set(CVPixelBufferRef image, size_t plane) noexcept {
-    unset();
-
+MetalExternalImage MetalExternalImage::createFromImagePlane(
+        MetalContext& context, CVPixelBufferRef image, uint32_t plane) {
     if (!image) {
-        return;
+        return {};
     }
 
     const OSType formatType = CVPixelBufferGetPixelFormatType(image);
     FILAMENT_CHECK_POSTCONDITION(formatType == kCVPixelFormatType_420YpCbCr8BiPlanarFullRange)
             << "Metal planar external images must be in the 420f format.";
+    FILAMENT_CHECK_POSTCONDITION(plane == 0 || plane == 1)
+            << "Metal planar external images must be created from planes 0 or 1.";
 
-    mImage = image;
-
-    auto getPlaneFormat = [] (size_t plane) {
-      // Right now Metal only supports kCVPixelFormatType_420YpCbCr8BiPlanarFullRange planar
-      // external images, so we can make the following assumptions about the format of each plane.
-      if (plane == 0) {
-        return MTLPixelFormatR8Unorm; // luminance
-      }
-      if (plane == 1) {
-        // CbCr
-        return MTLPixelFormatRG8Unorm; // CbCr
-      }
-      return MTLPixelFormatInvalid;
+    auto getPlaneFormat = [](size_t plane) {
+        // Right now Metal only supports kCVPixelFormatType_420YpCbCr8BiPlanarFullRange planar
+        // external images, so we can make the following assumptions about the format of each plane.
+        if (plane == 0) {
+            return MTLPixelFormatR8Unorm;  // luminance
+        }
+        if (plane == 1) {
+            return MTLPixelFormatRG8Unorm;  // CbCr
+        }
+        return MTLPixelFormatInvalid;
     };
 
     const MTLPixelFormat format = getPlaneFormat(plane);
     assert_invariant(format != MTLPixelFormatInvalid);
-    mTexture = createTextureFromImage(image, format, plane);
-    mTextureView = createSwizzledTextureView(mTexture);
+    CVMetalTextureRef mTexture = createTextureFromImage(context.textureCache, image, format, plane);
+    return { CVPixelBufferRetain(image), mTexture };
 }
 
-id<MTLTexture> MetalExternalImage::getMetalTextureForDraw() const noexcept {
+MetalExternalImage::MetalExternalImage(MetalExternalImage&& rhs) {
+    std::swap(mImage, rhs.mImage);
+    std::swap(mTexture, rhs.mTexture);
+    std::swap(mRgbTexture, rhs.mRgbTexture);
+}
+
+MetalExternalImage& MetalExternalImage::operator=(MetalExternalImage&& rhs) {
+    CVPixelBufferRelease(mImage);
+    CVBufferRelease(mTexture);
+    mImage = nullptr;
+    mTexture = nullptr;
+    mRgbTexture = nullptr;
+    std::swap(mImage, rhs.mImage);
+    std::swap(mTexture, rhs.mTexture);
+    std::swap(mRgbTexture, rhs.mRgbTexture);
+    return *this;
+}
+
+MetalExternalImage::~MetalExternalImage() noexcept {
+    CVPixelBufferRelease(mImage);
+    CVBufferRelease(mTexture);
+}
+
+id<MTLTexture> MetalExternalImage::getMtlTexture() const noexcept {
     if (mRgbTexture) {
         return mRgbTexture;
     }
-
-    // Retain the image and Metal texture until the GPU has finished with this frame. This does
-    // not need to be done for the RGB texture, because it is an Objective-C object whose
-    // lifetime is automatically managed by Metal.
-    auto& tracker = mContext.resourceTracker;
-    auto commandBuffer = getPendingCommandBuffer(&mContext);
-    if (tracker.trackResource((__bridge void*) commandBuffer, mImage, cvBufferDeleter)) {
-        CVPixelBufferRetain(mImage);
+    if (mTexture) {
+        return CVMetalTextureGetTexture(mTexture);
     }
-    if (tracker.trackResource((__bridge void*) commandBuffer, mTexture, cvBufferDeleter)) {
-        CVBufferRetain(mTexture);
-    }
-
-    assert_invariant(mTextureView);
-    return mTextureView;
+    return nil;
 }
 
-CVMetalTextureRef MetalExternalImage::createTextureFromImage(CVPixelBufferRef image,
-        MTLPixelFormat format, size_t plane) {
+CVMetalTextureRef MetalExternalImage::createTextureFromImage(CVMetalTextureCacheRef textureCache,
+        CVPixelBufferRef image, MTLPixelFormat format, size_t plane) {
     const size_t width = CVPixelBufferGetWidthOfPlane(image, plane);
     const size_t height = CVPixelBufferGetHeightOfPlane(image, plane);
 
     CVMetalTextureRef texture;
-    CVReturn result = CVMetalTextureCacheCreateTextureFromImage(kCFAllocatorDefault,
-            mContext.textureCache, image, nullptr, format, width, height, plane, &texture);
+    CVReturn result = CVMetalTextureCacheCreateTextureFromImage(kCFAllocatorDefault, textureCache,
+            image, nullptr, format, width, height, plane, &texture);
     FILAMENT_CHECK_POSTCONDITION(result == kCVReturnSuccess)
             << "Could not create a CVMetalTexture from CVPixelBuffer.";
 
@@ -201,58 +221,19 @@ void MetalExternalImage::shutdown(MetalContext& context) noexcept {
     context.externalImageComputePipelineState = nil;
 }
 
-void MetalExternalImage::assertWritableImage(CVPixelBufferRef image) {
-    OSType formatType = CVPixelBufferGetPixelFormatType(image);
-    FILAMENT_CHECK_PRECONDITION(formatType == kCVPixelFormatType_32BGRA)
-            << "Metal SwapChain images must be in the 32BGRA format.";
-}
-
-void MetalExternalImage::unset() {
-    CVPixelBufferRelease(mImage);
-    CVBufferRelease(mTexture);
-
-    mImage = nullptr;
-    mTexture = nullptr;
-    mTextureView = nil;
-    mRgbTexture = nil;
-    mWidth = 0;
-    mHeight = 0;
-}
-
-id<MTLTexture> MetalExternalImage::createRgbTexture(size_t width, size_t height) {
+id<MTLTexture> MetalExternalImage::createRgbTexture(
+        id<MTLDevice> device, size_t width, size_t height) {
     MTLTextureDescriptor *descriptor =
             [MTLTextureDescriptor texture2DDescriptorWithPixelFormat:MTLPixelFormatRGBA8Unorm
                                                                width:width
                                                               height:height
                                                            mipmapped:NO];
     descriptor.usage = MTLTextureUsageShaderWrite | MTLTextureUsageShaderRead;
-    return [mContext.device newTextureWithDescriptor:descriptor];
+    return [device newTextureWithDescriptor:descriptor];
 }
 
-id<MTLTexture> MetalExternalImage::createSwizzledTextureView(id<MTLTexture> texture) const {
-    const bool isDefaultSwizzle =
-            mSwizzle.r == TextureSwizzle::CHANNEL_0 &&
-            mSwizzle.g == TextureSwizzle::CHANNEL_1 &&
-            mSwizzle.b == TextureSwizzle::CHANNEL_2 &&
-            mSwizzle.a == TextureSwizzle::CHANNEL_3;
-    if (!isDefaultSwizzle && mContext.supportsTextureSwizzling) {
-        // Even though we've already checked supportsTextureSwizzling, we still need to guard these
-        // calls with @availability, otherwise the API usage will generate compiler warnings.
-        if (@available(iOS 13, *)) {
-            texture = createTextureViewWithSwizzle(texture,
-                    getSwizzleChannels(mSwizzle.r, mSwizzle.g, mSwizzle.b, mSwizzle.a));
-        }
-    }
-    return texture;
-}
-
-id<MTLTexture> MetalExternalImage::createSwizzledTextureView(CVMetalTextureRef ref) const {
-    id<MTLTexture> texture = CVMetalTextureGetTexture(ref);
-    return createSwizzledTextureView(texture);
-}
-
-void MetalExternalImage::ensureComputePipelineState() {
-    if (mContext.externalImageComputePipelineState != nil) {
+void MetalExternalImage::ensureComputePipelineState(MetalContext& context) {
+    if (context.externalImageComputePipelineState != nil) {
         return;
     }
 
@@ -260,29 +241,28 @@ void MetalExternalImage::ensureComputePipelineState() {
 
     NSString* objcSource = [NSString stringWithCString:kernel
                                               encoding:NSUTF8StringEncoding];
-    id<MTLLibrary> library = [mContext.device newLibraryWithSource:objcSource
-                                                            options:nil
-                                                              error:&error];
+    id<MTLLibrary> library = [context.device newLibraryWithSource:objcSource
+                                                          options:nil
+                                                            error:&error];
     NSERROR_CHECK("Unable to compile Metal shading library.");
 
     id<MTLFunction> kernelFunction = [library newFunctionWithName:@"ycbcrToRgb"];
 
-    mContext.externalImageComputePipelineState =
-            [mContext.device newComputePipelineStateWithFunction:kernelFunction
-                                                           error:&error];
+    context.externalImageComputePipelineState =
+            [context.device newComputePipelineStateWithFunction:kernelFunction error:&error];
     NSERROR_CHECK("Unable to create Metal compute pipeline state.");
 }
 
-id<MTLCommandBuffer> MetalExternalImage::encodeColorConversionPass(id<MTLTexture> inYPlane,
-        id<MTLTexture> inCbCrTexture, id<MTLTexture> outTexture) {
-    ensureComputePipelineState();
+id<MTLCommandBuffer> MetalExternalImage::encodeColorConversionPass(MetalContext& context,
+        id<MTLTexture> inYPlane, id<MTLTexture> inCbCrTexture, id<MTLTexture> outTexture) {
+    ensureComputePipelineState(context);
 
-    id<MTLCommandBuffer> commandBuffer = [mContext.commandQueue commandBuffer];
+    id<MTLCommandBuffer> commandBuffer = [context.commandQueue commandBuffer];
     commandBuffer.label = @"YCbCr to RGB conversion";
 
     id<MTLComputeCommandEncoder> computeEncoder = [commandBuffer computeCommandEncoder];
 
-    [computeEncoder setComputePipelineState:mContext.externalImageComputePipelineState];
+    [computeEncoder setComputePipelineState:context.externalImageComputePipelineState];
     [computeEncoder setTexture:inYPlane atIndex:0];
     [computeEncoder setTexture:inCbCrTexture atIndex:1];
     [computeEncoder setTexture:outTexture atIndex:2];
@@ -300,5 +280,11 @@ id<MTLCommandBuffer> MetalExternalImage::encodeColorConversionPass(id<MTLTexture
     return commandBuffer;
 }
 
+void MetalExternalImage::assertWritableImage(CVPixelBufferRef image) {
+    OSType formatType = CVPixelBufferGetPixelFormatType(image);
+    FILAMENT_CHECK_PRECONDITION(formatType == kCVPixelFormatType_32BGRA)
+            << "Metal SwapChain images must be in the 32BGRA format.";
+}
+
 } // namespace backend
 } // namespace filament

filament/backend/src/metal/MetalHandles.h

@@ -44,6 +44,7 @@
 #include <condition_variable>
 #include <memory>
 #include <type_traits>
+#include <vector>
 
 namespace filament {
 namespace backend {
@@ -73,7 +74,8 @@ public:
 
     void releaseDrawable();
 
-    void setFrameScheduledCallback(CallbackHandler* handler, FrameScheduledCallback&& callback);
+    void setFrameScheduledCallback(
+            CallbackHandler* handler, FrameScheduledCallback&& callback, uint64_t flags);
     void setFrameCompletedCallback(
             CallbackHandler* handler, utils::Invocable<void(void)>&& callback);
 
@@ -84,6 +86,8 @@ public:
     NSUInteger getSurfaceWidth() const;
     NSUInteger getSurfaceHeight() const;
 
+    bool isPixelBuffer() const { return type == SwapChainType::CVPIXELBUFFERREF; }
+
 private:
 
     enum class SwapChainType {
@@ -93,7 +97,6 @@ private:
     };
     bool isCaMetalLayer() const { return type == SwapChainType::CAMETALLAYER; }
     bool isHeadless() const { return type == SwapChainType::HEADLESS; }
-    bool isPixelBuffer() const { return type == SwapChainType::CVPIXELBUFFERREF; }
 
     void scheduleFrameScheduledCallback();
     void scheduleFrameCompletedCallback();
@@ -109,6 +112,7 @@ private:
     NSUInteger headlessWidth = 0;
     NSUInteger headlessHeight = 0;
     CAMetalLayer* layer = nullptr;
+    std::shared_ptr<std::mutex> layerDrawableMutex;
     MetalExternalImage externalImage;
     SwapChainType type;
 
@@ -120,6 +124,7 @@ private:
     struct {
         CallbackHandler* handler = nullptr;
         std::shared_ptr<FrameScheduledCallback> callback = nullptr;
+        uint64_t flags = 0;
     } frameScheduled;
 
     struct {
@@ -130,19 +135,17 @@ private:
 
 class MetalBufferObject : public HwBufferObject {
 public:
+
+    using TagResolver = MetalBuffer::TagResolver;
+
     MetalBufferObject(MetalContext& context, BufferObjectBinding bindingType, BufferUsage usage,
          uint32_t byteCount);
 
-    void updateBuffer(void* data, size_t size, uint32_t byteOffset);
-    void updateBufferUnsynchronized(void* data, size_t size, uint32_t byteOffset);
+    void updateBuffer(void* data, size_t size, uint32_t byteOffset, TagResolver&& getHandleTag);
+    void updateBufferUnsynchronized(
+            void* data, size_t size, uint32_t byteOffset, TagResolver&& getHandleTag);
     MetalBuffer* getBuffer() { return &buffer; }
 
-    // Tracks which uniform/ssbo buffers this buffer object is bound into.
-    static_assert(Program::UNIFORM_BINDING_COUNT <= 32);
-    static_assert(MAX_SSBO_COUNT <= 32);
-    utils::bitset32 boundUniformBuffers;
-    utils::bitset32 boundSsbos;
-
 private:
     MetalBuffer buffer;
 };
@@ -199,12 +202,10 @@ public:
     MetalProgram(MetalContext& context, Program&& program) noexcept;
 
     const MetalShaderCompiler::MetalFunctionBundle& getFunctions();
-    const Program::SamplerGroupInfo& getSamplerGroupInfo() { return samplerGroupInfo; }
 
 private:
     void initialize();
 
-    Program::SamplerGroupInfo samplerGroupInfo;
     MetalContext& mContext;
     MetalShaderCompiler::MetalFunctionBundle mFunctionBundle;
     MetalShaderCompiler::program_token_t mToken;
@@ -226,43 +227,42 @@ struct PixelBufferShape {
 class MetalTexture : public HwTexture {
 public:
     MetalTexture(MetalContext& context, SamplerType target, uint8_t levels, TextureFormat format,
-            uint8_t samples, uint32_t width, uint32_t height, uint32_t depth, TextureUsage usage,
-            TextureSwizzle r, TextureSwizzle g, TextureSwizzle b, TextureSwizzle a)
-            noexcept;
+            uint8_t samples, uint32_t width, uint32_t height, uint32_t depth,
+            TextureUsage usage) noexcept;
+
+    // constructors for creating texture views
+    MetalTexture(MetalContext& context, MetalTexture const* src, uint8_t baseLevel,
+            uint8_t levelCount) noexcept;
+    MetalTexture(MetalContext& context, MetalTexture const* src, TextureSwizzle r, TextureSwizzle g,
+            TextureSwizzle b, TextureSwizzle a) noexcept;
 
     // Constructor for importing an id<MTLTexture> outside of Filament.
     MetalTexture(MetalContext& context, SamplerType target, uint8_t levels, TextureFormat format,
             uint8_t samples, uint32_t width, uint32_t height, uint32_t depth, TextureUsage usage,
             id<MTLTexture> texture) noexcept;
 
-    ~MetalTexture();
+    // Constructors for importing external images.
+    MetalTexture(MetalContext& context, TextureFormat format, uint32_t width, uint32_t height,
+            TextureUsage usage, CVPixelBufferRef image) noexcept;
+    MetalTexture(MetalContext& context, TextureFormat format, uint32_t width, uint32_t height,
+            TextureUsage usage, CVPixelBufferRef image, uint32_t plane) noexcept;
 
-    // Returns an id<MTLTexture> suitable for reading in a shader, taking into account swizzle and
-    // LOD clamping.
-    id<MTLTexture> getMtlTextureForRead() noexcept;
+    // Returns an id<MTLTexture> suitable for reading in a shader, taking into account swizzle.
+    id<MTLTexture> getMtlTextureForRead() const noexcept;
 
     // Returns the id<MTLTexture> for attaching to a render pass.
-    id<MTLTexture> getMtlTextureForWrite() noexcept {
+    id<MTLTexture> getMtlTextureForWrite() const noexcept {
         return texture;
     }
 
+    std::shared_ptr<MetalExternalImage> getExternalImage() const noexcept { return externalImage; }
+
     void loadImage(uint32_t level, MTLRegion region, PixelBufferDescriptor& p) noexcept;
     void generateMipmaps() noexcept;
 
-    // A texture starts out with none of its mip levels (also referred to as LODs) available for
-    // reading. 4 actions update the range of LODs available:
-    // - calling loadImage
-    // - calling generateMipmaps
-    // - using the texture as a render target attachment
-    // - calling setMinMaxLevels
-    // A texture's available mips are consistent throughout a render pass.
-    void setLodRange(uint16_t minLevel, uint16_t maxLevel);
-    void extendLodRangeTo(uint16_t level);
-
     static MTLPixelFormat decidePixelFormat(MetalContext* context, TextureFormat format);
 
     MetalContext& context;
-    MetalExternalImage externalImage;
 
     // A "sidecar" texture used to implement automatic MSAA resolve.
     // This is created by MetalRenderTarget and stored here so it can be used with multiple
@@ -271,26 +271,6 @@ public:
 
     MTLPixelFormat devicePixelFormat;
 
-    // Frees memory associated with this texture and marks it as "terminated".
-    // Used to track "use after free" scenario.
-    void terminate() noexcept;
-    bool isTerminated() const noexcept { return terminated; }
-    inline void checkUseAfterFree(const char* samplerGroupDebugName, size_t textureIndex) const {
-        if (UTILS_LIKELY(!isTerminated())) {
-            return;
-        }
-        NSString* reason =
-                [NSString stringWithFormat:
-                                  @"Filament Metal texture use after free, sampler group = "
-                                  @"%s, texture index = %zu",
-                          samplerGroupDebugName, textureIndex];
-        NSException* useAfterFreeException =
-                [NSException exceptionWithName:@"MetalTextureUseAfterFree"
-                                        reason:reason
-                                      userInfo:nil];
-        [useAfterFreeException raise];
-    }
-
 private:
     void loadSlice(uint32_t level, MTLRegion region, uint32_t byteOffset, uint32_t slice,
             PixelBufferDescriptor const& data) noexcept;
@@ -301,95 +281,12 @@ private:
 
     id<MTLTexture> texture = nil;
 
+    std::shared_ptr<MetalExternalImage> externalImage;
+
     // If non-nil, a swizzled texture view to use instead of "texture".
     // Filament swizzling only affects texture reads, so this should not be used when the texture is
     // bound as a render target attachment.
     id<MTLTexture> swizzledTextureView = nil;
-    id<MTLTexture> lodTextureView = nil;
-
-    uint16_t minLod = std::numeric_limits<uint16_t>::max();
-    uint16_t maxLod = 0;
-
-    bool terminated = false;
-};
-
-class MetalSamplerGroup : public HwSamplerGroup {
-public:
-    explicit MetalSamplerGroup(size_t size, utils::FixedSizeString<32> name) noexcept
-        : size(size),
-          debugName(name),
-          textureHandles(size, Handle<HwTexture>()),
-          textures(size, nil),
-          samplers(size, nil) {}
-
-    inline void setTextureHandle(size_t index, Handle<HwTexture> th) {
-        assert_invariant(!finalized);
-        textureHandles[index] = th;
-    }
-
-    // This method is only used for debugging, to ensure all texture handles are alive.
-    const auto& getTextureHandles() const {
-        return textureHandles;
-    }
-
-    // Encode a MTLTexture into this SamplerGroup at the given index.
-    inline void setFinalizedTexture(size_t index, id<MTLTexture> t) {
-        assert_invariant(!finalized);
-        textures[index] = t;
-    }
-
-    // Encode a MTLSamplerState into this SamplerGroup at the given index.
-    inline void setFinalizedSampler(size_t index, id<MTLSamplerState> s) {
-        assert_invariant(!finalized);
-        samplers[index] = s;
-    }
-
-    // A SamplerGroup is "finalized" when all of its textures have been set and is ready for use in
-    // a draw call.
-    // Once a SamplerGroup is finalized, it must be reset or mutated to be written into again.
-    void finalize();
-    bool isFinalized() const noexcept { return finalized; }
-
-    // Both of these methods "unfinalize" a SamplerGroup, allowing it to be updated via calls to
-    // setFinalizedTexture or setFinalizedSampler. The difference is that when reset is called, all
-    // the samplers/textures must be rebound. The MTLArgumentEncoder must be specified, in case
-    // the texture types have changed.
-    // Mutate re-encodes the current set of samplers/textures into the new argument
-    // buffer.
-    void reset(id<MTLCommandBuffer> cmdBuffer, id<MTLArgumentEncoder> e, id<MTLDevice> device);
-    void mutate(id<MTLCommandBuffer> cmdBuffer);
-
-    id<MTLBuffer> getArgumentBuffer() const {
-        assert_invariant(finalized);
-        return argBuffer->getCurrentAllocation().first;
-    }
-
-    NSUInteger getArgumentBufferOffset() const {
-        return argBuffer->getCurrentAllocation().second;
-    }
-
-    inline std::pair<Handle<HwTexture>, id<MTLTexture>> getFinalizedTexture(size_t index) {
-        return {textureHandles[index], textures[index]};
-    }
-
-    // Calls the Metal useResource:usage:stages: method for all the textures in this SamplerGroup.
-    void useResources(id<MTLRenderCommandEncoder> renderPassEncoder);
-
-    size_t size;
-    utils::FixedSizeString<32> debugName;
-
-public:
-
-    // These vectors are kept in sync with one another.
-    utils::FixedCapacityVector<Handle<HwTexture>> textureHandles;
-    utils::FixedCapacityVector<id<MTLTexture>> textures;
-    utils::FixedCapacityVector<id<MTLSamplerState>> samplers;
-
-    id<MTLArgumentEncoder> encoder;
-
-    std::unique_ptr<MetalRingBuffer> argBuffer = nullptr;
-
-    bool finalized = false;
 };
 
 class MetalRenderTarget : public HwRenderTarget {
@@ -546,6 +443,61 @@ struct MetalTimerQuery : public HwTimerQuery {
     std::shared_ptr<Status> status;
 };
 
+class MetalDescriptorSetLayout : public HwDescriptorSetLayout {
+public:
+    MetalDescriptorSetLayout(DescriptorSetLayout&& layout) noexcept;
+
+    const auto& getBindings() const noexcept { return mLayout.bindings; }
+
+    size_t getDynamicOffsetCount() const noexcept { return mDynamicOffsetCount; }
+
+    /**
+     * Get an argument encoder for this descriptor set and shader stage.
+     * textureTypes should only include the textures present in the corresponding shader stage.
+     */
+    id<MTLArgumentEncoder> getArgumentEncoder(id<MTLDevice> device, ShaderStage stage,
+            utils::FixedCapacityVector<MTLTextureType> const& textureTypes);
+
+private:
+    id<MTLArgumentEncoder> getArgumentEncoderSlow(id<MTLDevice> device, ShaderStage stage,
+            utils::FixedCapacityVector<MTLTextureType> const& textureTypes);
+
+    DescriptorSetLayout mLayout;
+    size_t mDynamicOffsetCount = 0;
+    std::array<id<MTLArgumentEncoder>, Program::SHADER_TYPE_COUNT> mCachedArgumentEncoder = { nil };
+    std::array<utils::FixedCapacityVector<MTLTextureType>, Program::SHADER_TYPE_COUNT>
+            mCachedTextureTypes;
+};
+
+struct MetalDescriptorSet : public HwDescriptorSet {
+    MetalDescriptorSet(MetalDescriptorSetLayout* layout) noexcept;
+
+    void finalize(MetalDriver* driver);
+
+    id<MTLBuffer> finalizeAndGetBuffer(MetalDriver* driver, ShaderStage stage);
+
+    MetalDescriptorSetLayout* layout;
+
+    struct BufferBinding {
+        id<MTLBuffer> buffer;
+        uint32_t offset;
+        uint32_t size;
+    };
+    struct TextureBinding {
+        id<MTLTexture> texture;
+        SamplerParams sampler;
+    };
+    tsl::robin_map<descriptor_binding_t, BufferBinding> buffers;
+    tsl::robin_map<descriptor_binding_t, TextureBinding> textures;
+
+    std::vector<id<MTLResource>> vertexResources;
+    std::vector<id<MTLResource>> fragmentResources;
+
+    std::vector<std::shared_ptr<MetalExternalImage>> externalImages;
+
+    std::array<TrackedMetalBuffer, Program::SHADER_TYPE_COUNT> cachedBuffer = { nil };
+};
+
 } // namespace backend
 } // namespace filament
 

filament/backend/src/metal/MetalHandles.mm

@@ -73,7 +73,7 @@ MetalSwapChain::MetalSwapChain(MetalContext& context, CAMetalLayer* nativeWindow
     : context(context),
       depthStencilFormat(decideDepthStencilFormat(flags)),
       layer(nativeWindow),
-      externalImage(context),
+      layerDrawableMutex(std::make_shared<std::mutex>()),
       type(SwapChainType::CAMETALLAYER) {
 
     if (!(flags & SwapChain::CONFIG_TRANSPARENT) && !nativeWindow.opaque) {
@@ -99,17 +99,15 @@ MetalSwapChain::MetalSwapChain(MetalContext& context, int32_t width, int32_t hei
       depthStencilFormat(decideDepthStencilFormat(flags)),
       headlessWidth(width),
       headlessHeight(height),
-      externalImage(context),
       type(SwapChainType::HEADLESS) {}
 
 MetalSwapChain::MetalSwapChain(MetalContext& context, CVPixelBufferRef pixelBuffer, uint64_t flags)
     : context(context),
       depthStencilFormat(decideDepthStencilFormat(flags)),
-      externalImage(context),
+      externalImage(MetalExternalImage::createFromImage(context, pixelBuffer)),
       type(SwapChainType::CVPIXELBUFFERREF) {
     assert_invariant(flags & SWAP_CHAIN_CONFIG_APPLE_CVPIXELBUFFER);
     MetalExternalImage::assertWritableImage(pixelBuffer);
-    externalImage.set(pixelBuffer);
     assert_invariant(externalImage.isValid());
 }
 
@@ -120,7 +118,6 @@ MTLPixelFormat MetalSwapChain::decideDepthStencilFormat(uint64_t flags) {
 }
 
 MetalSwapChain::~MetalSwapChain() {
-    externalImage.set(nullptr);
 }
 
 NSUInteger MetalSwapChain::getSurfaceWidth() const {
@@ -170,18 +167,28 @@ id<MTLTexture> MetalSwapChain::acquireDrawable() {
     }
 
     if (isPixelBuffer()) {
-        return externalImage.getMetalTextureForDraw();
+        return externalImage.getMtlTexture();
     }
 
     assert_invariant(isCaMetalLayer());
-    drawable = [layer nextDrawable];
+
+    // CAMetalLayer's drawable pool is not thread safe. Use a mutex when
+    // calling -nextDrawable, or when releasing the last known reference
+    // to any CAMetalDrawable returned from a previous -nextDrawable.
+    {
+        std::lock_guard<std::mutex> lock(*layerDrawableMutex);
+        drawable = [layer nextDrawable];
+    }
 
     FILAMENT_CHECK_POSTCONDITION(drawable != nil) << "Could not obtain drawable.";
     return drawable.texture;
 }
 
 void MetalSwapChain::releaseDrawable() {
-    drawable = nil;
+    if (drawable) {
+        std::lock_guard<std::mutex> lock(*layerDrawableMutex);
+        drawable = nil;
+    }
 }
 
 id<MTLTexture> MetalSwapChain::acquireDepthTexture() {
@@ -222,9 +229,10 @@ void MetalSwapChain::ensureDepthStencilTexture() {
 }
 
 void MetalSwapChain::setFrameScheduledCallback(
-        CallbackHandler* handler, FrameScheduledCallback&& callback) {
+        CallbackHandler* handler, FrameScheduledCallback&& callback, uint64_t flags) {
     frameScheduled.handler = handler;
     frameScheduled.callback = std::make_shared<FrameScheduledCallback>(std::move(callback));
+    frameScheduled.flags = flags;
 }
 
 void MetalSwapChain::setFrameCompletedCallback(
@@ -246,19 +254,17 @@ void MetalSwapChain::present() {
     }
 }
 
-#ifndef FILAMENT_RELEASE_PRESENT_DRAWABLE_MAIN_THREAD
-#define FILAMENT_RELEASE_PRESENT_DRAWABLE_MAIN_THREAD 1
-#endif
-
 class PresentDrawableData {
 public:
     PresentDrawableData() = delete;
     PresentDrawableData(const PresentDrawableData&) = delete;
     PresentDrawableData& operator=(const PresentDrawableData&) = delete;
 
-    static PresentDrawableData* create(id<CAMetalDrawable> drawable, MetalDriver* driver) {
+    static PresentDrawableData* create(id<CAMetalDrawable> drawable,
+            std::shared_ptr<std::mutex> drawableMutex, MetalDriver* driver, uint64_t flags) {
+        assert_invariant(drawableMutex);
         assert_invariant(driver);
-        return new PresentDrawableData(drawable, driver);
+        return new PresentDrawableData(drawable, drawableMutex, driver, flags);
     }
 
     static void maybePresentAndDestroyAsync(PresentDrawableData* that, bool shouldPresent) {
@@ -266,28 +272,38 @@ public:
            [that->mDrawable present];
         }
 
-#if FILAMENT_RELEASE_PRESENT_DRAWABLE_MAIN_THREAD == 1
-        // mDrawable is acquired on the driver thread. Typically, we would release this object on
-        // the same thread, but after receiving consistent crash reports from within
-        // [CAMetalDrawable dealloc], we suspect this object requires releasing on the main thread.
-        dispatch_async(dispatch_get_main_queue(), ^{ cleanupAndDestroy(that); });
-#else
-        that->mDriver->runAtNextTick([that]() { cleanupAndDestroy(that); });
-#endif
+        if (that->mFlags & SwapChain::CALLBACK_DEFAULT_USE_METAL_COMPLETION_HANDLER) {
+            cleanupAndDestroy(that);
+        } else {
+            // mDrawable is acquired on the driver thread. Typically, we would release this object
+            // on the same thread, but after receiving consistent crash reports from within
+            // [CAMetalDrawable dealloc], we suspect this object requires releasing on the main
+            // thread.
+            dispatch_async(dispatch_get_main_queue(), ^{
+                cleanupAndDestroy(that);
+            });
+        }
     }
 
 private:
-    PresentDrawableData(id<CAMetalDrawable> drawable, MetalDriver* driver)
-        : mDrawable(drawable), mDriver(driver) {}
+    PresentDrawableData(id<CAMetalDrawable> drawable, std::shared_ptr<std::mutex> drawableMutex,
+            MetalDriver* driver, uint64_t flags)
+        : mDrawable(drawable), mDrawableMutex(drawableMutex), mDriver(driver), mFlags(flags) {}
 
     static void cleanupAndDestroy(PresentDrawableData *that) {
-        that->mDrawable = nil;
+        if (that->mDrawable) {
+            std::lock_guard<std::mutex> lock(*(that->mDrawableMutex));
+            that->mDrawable = nil;
+        }
+        that->mDrawableMutex.reset();
         that->mDriver = nullptr;
         delete that;
     }
 
     id<CAMetalDrawable> mDrawable;
+    std::shared_ptr<std::mutex> mDrawableMutex;
     MetalDriver* mDriver = nullptr;
+    uint64_t mFlags = 0;
 };
 
 void presentDrawable(bool presentFrame, void* user) {
@@ -304,8 +320,8 @@ void MetalSwapChain::scheduleFrameScheduledCallback() {
 
     struct Callback {
         Callback(std::shared_ptr<FrameScheduledCallback> callback, id<CAMetalDrawable> drawable,
-                MetalDriver* driver)
-            : f(callback), data(PresentDrawableData::create(drawable, driver)) {}
+                 std::shared_ptr<std::mutex> drawableMutex, MetalDriver* driver, uint64_t flags)
+            : f(callback), data(PresentDrawableData::create(drawable, drawableMutex, driver, flags)) {}
         std::shared_ptr<FrameScheduledCallback> f;
         // PresentDrawableData* is destroyed by maybePresentAndDestroyAsync() later.
         std::unique_ptr<PresentDrawableData> data;
@@ -320,14 +336,19 @@ void MetalSwapChain::scheduleFrameScheduledCallback() {
 
     // This callback pointer will be captured by the block. Even if the scheduled handler is never
     // called, the unique_ptr will still ensure we don't leak memory.
-    __block auto callback =
-            std::make_unique<Callback>(frameScheduled.callback, drawable, context.driver);
+    uint64_t const flags = frameScheduled.flags;
+    __block auto callback = std::make_unique<Callback>(
+            frameScheduled.callback, drawable, layerDrawableMutex, context.driver, flags);
 
     backend::CallbackHandler* handler = frameScheduled.handler;
     MetalDriver* driver = context.driver;
     [getPendingCommandBuffer(&context) addScheduledHandler:^(id<MTLCommandBuffer> cb) {
         Callback* user = callback.release();
-        driver->scheduleCallback(handler, user, &Callback::func);
+        if (flags & SwapChain::CALLBACK_DEFAULT_USE_METAL_COMPLETION_HANDLER) {
+            Callback::func(user);
+        } else {
+            driver->scheduleCallback(handler, user, &Callback::func);
+        }
     }];
 }
 
@@ -362,12 +383,14 @@ MetalBufferObject::MetalBufferObject(MetalContext& context, BufferObjectBinding
         BufferUsage usage, uint32_t byteCount)
         : HwBufferObject(byteCount), buffer(context, bindingType, usage, byteCount) {}
 
-void MetalBufferObject::updateBuffer(void* data, size_t size, uint32_t byteOffset) {
-    buffer.copyIntoBuffer(data, size, byteOffset);
+void MetalBufferObject::updateBuffer(
+        void* data, size_t size, uint32_t byteOffset, TagResolver&& getHandleTag) {
+    buffer.copyIntoBuffer(data, size, byteOffset, std::move(getHandleTag));
 }
 
-void MetalBufferObject::updateBufferUnsynchronized(void* data, size_t size, uint32_t byteOffset) {
-    buffer.copyIntoBufferUnsynchronized(data, size, byteOffset);
+void MetalBufferObject::updateBufferUnsynchronized(
+        void* data, size_t size, uint32_t byteOffset, TagResolver&& getHandleTag) {
+    buffer.copyIntoBufferUnsynchronized(data, size, byteOffset, std::move(getHandleTag));
 }
 
 MetalVertexBufferInfo::MetalVertexBufferInfo(MetalContext& context, uint8_t bufferCount,
@@ -463,11 +486,6 @@ void MetalRenderPrimitive::setBuffers(MetalVertexBufferInfo const* const vbi,
 
 MetalProgram::MetalProgram(MetalContext& context, Program&& program) noexcept
     : HwProgram(program.getName()), mContext(context) {
-
-    // Save this program's SamplerGroupInfo, it's used during draw calls to bind sampler groups to
-    // the appropriate stage(s).
-    samplerGroupInfo = program.getSamplerGroupInfo();
-
     mToken = context.shaderCompiler->createProgram(program.getName(), std::move(program));
     assert_invariant(mToken);
 }
@@ -487,10 +505,9 @@ void MetalProgram::initialize() {
 
 MetalTexture::MetalTexture(MetalContext& context, SamplerType target, uint8_t levels,
         TextureFormat format, uint8_t samples, uint32_t width, uint32_t height, uint32_t depth,
-        TextureUsage usage, TextureSwizzle r, TextureSwizzle g, TextureSwizzle b,
-        TextureSwizzle a) noexcept
-    : HwTexture(target, levels, samples, width, height, depth, format, usage), context(context),
-        externalImage(context, r, g, b, a) {
+        TextureUsage usage) noexcept
+    : HwTexture(target, levels, samples, width, height, depth, format, usage), context(context) {
+    assert_invariant(target != SamplerType::SAMPLER_EXTERNAL);
 
     devicePixelFormat = decidePixelFormat(&context, format);
     FILAMENT_CHECK_POSTCONDITION(devicePixelFormat != MTLPixelFormatInvalid)
@@ -537,8 +554,6 @@ MetalTexture::MetalTexture(MetalContext& context, SamplerType target, uint8_t le
             descriptor.usage = getMetalTextureUsage(usage);
             descriptor.storageMode = MTLStorageModePrivate;
             texture = [context.device newTextureWithDescriptor:descriptor];
-            FILAMENT_CHECK_POSTCONDITION(texture != nil)
-                    << "Could not create Metal texture. Out of memory?";
             break;
         case SamplerType::SAMPLER_CUBEMAP:
         case SamplerType::SAMPLER_CUBEMAP_ARRAY:
@@ -553,8 +568,6 @@ MetalTexture::MetalTexture(MetalContext& context, SamplerType target, uint8_t le
             descriptor.usage = getMetalTextureUsage(usage);
             descriptor.storageMode = MTLStorageModePrivate;
             texture = [context.device newTextureWithDescriptor:descriptor];
-            FILAMENT_CHECK_POSTCONDITION(texture != nil)
-                    << "Could not create Metal texture. Out of memory?";
             break;
         case SamplerType::SAMPLER_3D:
             descriptor = [MTLTextureDescriptor new];
@@ -567,8 +580,6 @@ MetalTexture::MetalTexture(MetalContext& context, SamplerType target, uint8_t le
             descriptor.usage = getMetalTextureUsage(usage);
             descriptor.storageMode = MTLStorageModePrivate;
             texture = [context.device newTextureWithDescriptor:descriptor];
-            FILAMENT_CHECK_POSTCONDITION(texture != nil)
-                    << "Could not create Metal texture. Out of memory?";
             break;
         case SamplerType::SAMPLER_EXTERNAL:
             // If we're using external textures (CVPixelBufferRefs), we don't need to make any
@@ -577,15 +588,33 @@ MetalTexture::MetalTexture(MetalContext& context, SamplerType target, uint8_t le
             break;
     }
 
-    // If swizzling is set, set up a swizzled texture view that we'll use when sampling this texture.
-    const bool isDefaultSwizzle =
-            r == TextureSwizzle::CHANNEL_0 &&
-            g == TextureSwizzle::CHANNEL_1 &&
-            b == TextureSwizzle::CHANNEL_2 &&
-            a == TextureSwizzle::CHANNEL_3;
-    // If texture is nil, then it must be a SAMPLER_EXTERNAL texture.
-    // Swizzling for external textures is handled inside MetalExternalImage.
-    if (!isDefaultSwizzle && texture && context.supportsTextureSwizzling) {
+    FILAMENT_CHECK_POSTCONDITION(target == SamplerType::SAMPLER_EXTERNAL || texture != nil)
+            << "Could not create Metal texture (SamplerType = " << int(target)
+            << ", levels = " << int(levels) << ", MTLPixelFormat = " << int(devicePixelFormat)
+            << ", width = " << width << ", height = " << height << ", depth = " << depth
+            << "). Out of memory?";
+}
+
+MetalTexture::MetalTexture(MetalContext& context, MetalTexture const* src, uint8_t baseLevel,
+        uint8_t levelCount) noexcept
+    : HwTexture(src->target, src->levels, src->samples, src->width, src->height, src->depth,
+              src->format, src->usage),
+      context(context),
+      devicePixelFormat(src->devicePixelFormat),
+      externalImage(src->externalImage) {
+    texture = createTextureViewWithLodRange(
+            src->getMtlTextureForRead(), baseLevel, baseLevel + levelCount - 1);
+}
+
+MetalTexture::MetalTexture(MetalContext& context, MetalTexture const* src, TextureSwizzle r,
+        TextureSwizzle g, TextureSwizzle b, TextureSwizzle a) noexcept
+    : HwTexture(src->target, src->levels, src->samples, src->width, src->height, src->depth,
+              src->format, src->usage),
+      context(context),
+      devicePixelFormat(src->devicePixelFormat),
+      externalImage(src->externalImage) {
+    texture = src->getMtlTextureForRead();
+    if (context.supportsTextureSwizzling) {
         // Even though we've already checked context.supportsTextureSwizzling, we still need to
         // guard these calls with @availability, otherwise the API usage will generate compiler
         // warnings.
@@ -599,44 +628,30 @@ MetalTexture::MetalTexture(MetalContext& context, SamplerType target, uint8_t le
 MetalTexture::MetalTexture(MetalContext& context, SamplerType target, uint8_t levels, TextureFormat format,
         uint8_t samples, uint32_t width, uint32_t height, uint32_t depth, TextureUsage usage,
         id<MTLTexture> metalTexture) noexcept
-    : HwTexture(target, levels, samples, width, height, depth, format, usage), context(context),
-        externalImage(context) {
+    : HwTexture(target, levels, samples, width, height, depth, format, usage), context(context) {
     texture = metalTexture;
-    setLodRange(0, levels - 1);
 }
 
-void MetalTexture::terminate() noexcept {
-    texture = nil;
-    swizzledTextureView = nil;
-    lodTextureView = nil;
-    msaaSidecar = nil;
-    externalImage.set(nullptr);
-    terminated = true;
+MetalTexture::MetalTexture(MetalContext& context, TextureFormat format, uint32_t width,
+        uint32_t height, TextureUsage usage, CVPixelBufferRef image) noexcept
+    : HwTexture(SamplerType::SAMPLER_EXTERNAL, 1, 1, width, height, 1, format, usage),
+      context(context),
+      externalImage(std::make_shared<MetalExternalImage>(
+              MetalExternalImage::createFromImage(context, image))) {
+    texture = externalImage->getMtlTexture();
 }
 
-MetalTexture::~MetalTexture() {
-    externalImage.set(nullptr);
+MetalTexture::MetalTexture(MetalContext& context, TextureFormat format, uint32_t width,
+        uint32_t height, TextureUsage usage, CVPixelBufferRef image, uint32_t plane) noexcept
+    : HwTexture(SamplerType::SAMPLER_EXTERNAL, 1, 1, width, height, 1, format, usage),
+      context(context),
+      externalImage(std::make_shared<MetalExternalImage>(
+              MetalExternalImage::createFromImagePlane(context, image, plane))) {
+    texture = externalImage->getMtlTexture();
 }
 
-id<MTLTexture> MetalTexture::getMtlTextureForRead() noexcept {
-    if (lodTextureView) {
-        return lodTextureView;
-    }
-    // The texture's swizzle remains constant throughout its lifetime, however its LOD range can
-    // change. We'll cache the LOD view, and set lodTextureView to nil if minLod or maxLod is
-    // updated.
-    id<MTLTexture> t = swizzledTextureView ? swizzledTextureView : texture;
-    if (!t) {
-        return nil;
-    }
-    if (UTILS_UNLIKELY(minLod > maxLod)) {
-        // If the texture does not have any available LODs, provide a view of only level 0.
-        // Filament should prevent this from ever occurring.
-        lodTextureView = createTextureViewWithLodRange(t, 0, 0);
-        return lodTextureView;
-    }
-    lodTextureView = createTextureViewWithLodRange(t, minLod, maxLod);
-    return lodTextureView;
+id<MTLTexture> MetalTexture::getMtlTextureForRead() const noexcept {
+    return swizzledTextureView ? swizzledTextureView : texture;
 }
 
 MTLPixelFormat MetalTexture::decidePixelFormat(MetalContext* context, TextureFormat format) {
@@ -755,15 +770,12 @@ void MetalTexture::loadImage(uint32_t level, MTLRegion region, PixelBufferDescri
             assert_invariant(false);
         }
     }
-
-    extendLodRangeTo(level);
 }
 
 void MetalTexture::generateMipmaps() noexcept {
     id <MTLBlitCommandEncoder> blitEncoder = [getPendingCommandBuffer(&context) blitCommandEncoder];
     [blitEncoder generateMipmapsForTexture:texture];
     [blitEncoder endEncoding];
-    setLodRange(0, texture.mipmapLevelCount - 1);
 }
 
 void MetalTexture::loadSlice(uint32_t level, MTLRegion region, uint32_t byteOffset, uint32_t slice,
@@ -887,98 +899,6 @@ void MetalTexture::loadWithBlit(uint32_t level, uint32_t slice, MTLRegion region
     context.blitter->blit(getPendingCommandBuffer(&context), args, "Texture upload blit");
 }
 
-void MetalTexture::extendLodRangeTo(uint16_t level) {
-    assert_invariant(!isInRenderPass(&context));
-    minLod = std::min(minLod, level);
-    maxLod = std::max(maxLod, level);
-    lodTextureView = nil;
-}
-
-void MetalTexture::setLodRange(uint16_t min, uint16_t max) {
-    assert_invariant(!isInRenderPass(&context));
-    assert_invariant(min <= max);
-    minLod = min;
-    maxLod = max;
-    lodTextureView = nil;
-}
-
-void MetalSamplerGroup::finalize() {
-    assert_invariant(encoder);
-    // TODO: we should be able to encode textures and samplers inside setFinalizedTexture and
-    // setFinalizedSampler as they become available, but Metal doesn't seem to like this; the arg
-    // buffer gets encoded incorrectly. This warrants more investigation.
-
-    auto [buffer, offset] = argBuffer->getCurrentAllocation();
-    [encoder setArgumentBuffer:buffer offset:offset];
-
-    // Encode all textures and samplers.
-    for (size_t s = 0; s < size; s++) {
-        [encoder setTexture:textures[s] atIndex:(s * 2 + 0)];
-        [encoder setSamplerState:samplers[s] atIndex:(s * 2 + 1)];
-    }
-
-    finalized = true;
-}
-
-void MetalSamplerGroup::reset(id<MTLCommandBuffer> cmdBuffer, id<MTLArgumentEncoder> e,
-        id<MTLDevice> device) {
-    encoder = e;
-
-    // The number of slots in the ring buffer we use to manage argument buffer allocations.
-    // This number was chosen to avoid running out of slots and having to allocate a "fallback"
-    // buffer when SamplerGroups are updated multiple times a frame. This value can reduced after
-    // auditing Filament's calls to updateSamplerGroup, which should be as few times as possible.
-    // For example, the bloom downsample pass should be refactored to maintain two separate
-    // MaterialInstances instead of "ping ponging" between two texture bindings, which causes a
-    // single SamplerGroup to be updated many times a frame.
-    static constexpr auto METAL_ARGUMENT_BUFFER_SLOTS = 32;
-
-    MTLSizeAndAlign argBufferLayout;
-    argBufferLayout.size = encoder.encodedLength;
-    argBufferLayout.align = encoder.alignment;
-    // Chances are, even though the MTLArgumentEncoder might change, the required size and alignment
-    // probably won't. So we can re-use the previous ring buffer.
-    if (UTILS_UNLIKELY(!argBuffer || !argBuffer->canAccomodateLayout(argBufferLayout))) {
-        argBuffer = std::make_unique<MetalRingBuffer>(device, MTLResourceStorageModeShared,
-                argBufferLayout, METAL_ARGUMENT_BUFFER_SLOTS);
-    } else {
-        argBuffer->createNewAllocation(cmdBuffer);
-    }
-
-    // Clear all textures and samplers.
-    assert_invariant(textureHandles.size() == textures.size());
-    assert_invariant(textures.size() == samplers.size());
-    for (size_t s = 0; s < textureHandles.size(); s++) {
-        textureHandles[s] = {};
-        textures[s] = nil;
-        samplers[s] = nil;
-    }
-
-    finalized = false;
-}
-
-void MetalSamplerGroup::mutate(id<MTLCommandBuffer> cmdBuffer) {
-    assert_invariant(finalized);    // only makes sense to mutate if this sampler group is finalized
-    assert_invariant(argBuffer);
-    argBuffer->createNewAllocation(cmdBuffer);
-    finalized = false;
-}
-
-void MetalSamplerGroup::useResources(id<MTLRenderCommandEncoder> renderPassEncoder) {
-    assert_invariant(finalized);
-    if (@available(iOS 13, *)) {
-        // TODO: pass only the appropriate stages to useResources.
-        [renderPassEncoder useResources:textures.data()
-                                  count:textures.size()
-                                  usage:MTLResourceUsageRead | MTLResourceUsageSample
-                                 stages:MTLRenderStageFragment | MTLRenderStageVertex];
-    } else {
-        [renderPassEncoder useResources:textures.data()
-                                  count:textures.size()
-                                  usage:MTLResourceUsageRead | MTLResourceUsageSample];
-    }
-}
-
 MetalRenderTarget::MetalRenderTarget(MetalContext* context, uint32_t width, uint32_t height,
         uint8_t samples, Attachment colorAttachments[MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT],
         Attachment depthAttachment, Attachment stencilAttachment) :
@@ -1324,5 +1244,193 @@ FenceStatus MetalFence::wait(uint64_t timeoutNs) {
     return FenceStatus::ERROR;
 }
 
+MetalDescriptorSetLayout::MetalDescriptorSetLayout(DescriptorSetLayout&& l) noexcept
+    : mLayout(std::move(l)) {
+    size_t dynamicBindings = 0;
+    for (const auto& binding : mLayout.bindings) {
+        if (any(binding.flags & DescriptorFlags::DYNAMIC_OFFSET)) {
+            dynamicBindings++;
+        }
+    }
+    mDynamicOffsetCount = dynamicBindings;
+}
+
+id<MTLArgumentEncoder> MetalDescriptorSetLayout::getArgumentEncoder(id<MTLDevice> device, ShaderStage stage,
+        utils::FixedCapacityVector<MTLTextureType> const& textureTypes) {
+    auto const index = static_cast<size_t>(stage);
+    assert_invariant(index < mCachedArgumentEncoder.size());
+    if (mCachedArgumentEncoder[index] &&
+            std::equal(
+                    textureTypes.begin(), textureTypes.end(), mCachedTextureTypes[index].begin())) {
+        return mCachedArgumentEncoder[index];
+    }
+    mCachedArgumentEncoder[index] = getArgumentEncoderSlow(device, stage, textureTypes);
+    mCachedTextureTypes[index] = textureTypes;
+    return mCachedArgumentEncoder[index];
+}
+
+id<MTLArgumentEncoder> MetalDescriptorSetLayout::getArgumentEncoderSlow(id<MTLDevice> device,
+        ShaderStage stage, utils::FixedCapacityVector<MTLTextureType> const& textureTypes) {
+    auto const& bindings = getBindings();
+    NSMutableArray<MTLArgumentDescriptor*>* arguments = [NSMutableArray new];
+    // Important! The bindings must be sorted by binding number. This has already been done inside
+    // createDescriptorSetLayout.
+    size_t textureIndex = 0;
+    for (auto const& binding : bindings) {
+        if (!hasShaderType(binding.stageFlags, stage)) {
+            continue;
+        }
+        switch (binding.type) {
+            case DescriptorType::UNIFORM_BUFFER:
+            case DescriptorType::SHADER_STORAGE_BUFFER: {
+                MTLArgumentDescriptor* bufferArgument = [MTLArgumentDescriptor argumentDescriptor];
+                bufferArgument.index = binding.binding * 2;
+                bufferArgument.dataType = MTLDataTypePointer;
+                bufferArgument.access = MTLArgumentAccessReadOnly;
+                [arguments addObject:bufferArgument];
+                break;
+            }
+            case DescriptorType::SAMPLER: {
+                MTLArgumentDescriptor* textureArgument = [MTLArgumentDescriptor argumentDescriptor];
+                textureArgument.index = binding.binding * 2;
+                textureArgument.dataType = MTLDataTypeTexture;
+                MTLTextureType textureType = MTLTextureType2D;
+                if (textureIndex < textureTypes.size()) {
+                    textureType = textureTypes[textureIndex++];
+                }
+                textureArgument.textureType = textureType;
+                textureArgument.access = MTLArgumentAccessReadOnly;
+                [arguments addObject:textureArgument];
+
+                MTLArgumentDescriptor* samplerArgument = [MTLArgumentDescriptor argumentDescriptor];
+                samplerArgument.index = binding.binding * 2 + 1;
+                samplerArgument.dataType = MTLDataTypeSampler;
+                textureArgument.access = MTLArgumentAccessReadOnly;
+                [arguments addObject:samplerArgument];
+                break;
+            }
+            case DescriptorType::INPUT_ATTACHMENT:
+                // TODO: support INPUT_ATTACHMENT
+                assert_invariant(false);
+                break;
+        }
+    }
+    return [device newArgumentEncoderWithArguments:arguments];
+}
+
+MetalDescriptorSet::MetalDescriptorSet(MetalDescriptorSetLayout* layout) noexcept
+    : layout(layout) {}
+
+void MetalDescriptorSet::finalize(MetalDriver* driver) {
+    [driver->mContext->currentRenderPassEncoder useResource:driver->mContext->emptyBuffer
+                                                      usage:MTLResourceUsageRead];
+    [driver->mContext->currentRenderPassEncoder
+            useResource:getOrCreateEmptyTexture(driver->mContext)
+                  usage:MTLResourceUsageRead];
+
+    if (@available(iOS 13.0, *)) {
+        [driver->mContext->currentRenderPassEncoder useResources:vertexResources.data()
+                                                           count:vertexResources.size()
+                                                           usage:MTLResourceUsageRead
+                                                          stages:MTLRenderStageVertex];
+        [driver->mContext->currentRenderPassEncoder useResources:fragmentResources.data()
+                                                           count:fragmentResources.size()
+                                                           usage:MTLResourceUsageRead
+                                                          stages:MTLRenderStageFragment];
+    } else {
+        [driver->mContext->currentRenderPassEncoder useResources:vertexResources.data()
+                                                           count:vertexResources.size()
+                                                           usage:MTLResourceUsageRead];
+        [driver->mContext->currentRenderPassEncoder useResources:fragmentResources.data()
+                                                           count:fragmentResources.size()
+                                                           usage:MTLResourceUsageRead];
+    }
+}
+
+id<MTLBuffer> MetalDescriptorSet::finalizeAndGetBuffer(MetalDriver* driver, ShaderStage stage) {
+    auto const index = static_cast<size_t>(stage);
+    assert_invariant(index < cachedBuffer.size());
+    auto& buffer = cachedBuffer[index];
+
+    if (buffer) {
+        return buffer.get();
+    }
+
+    // Map all the texture bindings to their respective texture types.
+    auto const& bindings = layout->getBindings();
+    auto textureTypes = utils::FixedCapacityVector<MTLTextureType>::with_capacity(bindings.size());
+    for (auto const& binding : bindings) {
+        if (!hasShaderType(binding.stageFlags, stage)) {
+            continue;
+        }
+        MTLTextureType textureType = MTLTextureType2D;
+        if (auto found = textures.find(binding.binding); found != textures.end()) {
+            auto const& textureBinding = textures[binding.binding];
+            textureType = textureBinding.texture.textureType;
+        }
+        textureTypes.push_back(textureType);
+    }
+
+    MetalContext const& context = *driver->mContext;
+
+    id<MTLArgumentEncoder> encoder =
+            layout->getArgumentEncoder(context.device, stage, textureTypes);
+
+    {
+        ScopedAllocationTimer timer("descriptor_set");
+        buffer = { [context.device newBufferWithLength:encoder.encodedLength
+                                               options:MTLResourceStorageModeShared],
+            TrackedMetalBuffer::Type::DESCRIPTOR_SET };
+    }
+    [encoder setArgumentBuffer:buffer.get() offset:0];
+
+    for (auto const& binding : bindings) {
+        if (!hasShaderType(binding.stageFlags, stage)) {
+            continue;
+        }
+        switch (binding.type) {
+            case DescriptorType::UNIFORM_BUFFER:
+            case DescriptorType::SHADER_STORAGE_BUFFER: {
+                auto found = buffers.find(binding.binding);
+                if (found == buffers.end()) {
+                    [encoder setBuffer:driver->mContext->emptyBuffer
+                                offset:0
+                               atIndex:binding.binding * 2];
+                    continue;
+                }
+
+                auto const& bufferBinding = buffers[binding.binding];
+                [encoder setBuffer:bufferBinding.buffer
+                            offset:bufferBinding.offset
+                           atIndex:binding.binding * 2];
+                break;
+            }
+            case DescriptorType::SAMPLER: {
+                auto found = textures.find(binding.binding);
+                if (found == textures.end()) {
+                    [encoder setTexture:driver->mContext->emptyTexture atIndex:binding.binding * 2];
+                    id<MTLSamplerState> sampler =
+                            driver->mContext->samplerStateCache.getOrCreateState({});
+                    [encoder setSamplerState:sampler atIndex:binding.binding * 2 + 1];
+                    continue;
+                }
+
+                auto const& textureBinding = textures[binding.binding];
+                [encoder setTexture:textureBinding.texture atIndex:binding.binding * 2];
+                SamplerState samplerState { .samplerParams = textureBinding.sampler };
+                id<MTLSamplerState> sampler =
+                        driver->mContext->samplerStateCache.getOrCreateState(samplerState);
+                [encoder setSamplerState:sampler atIndex:binding.binding * 2 + 1];
+                break;
+            }
+            case DescriptorType::INPUT_ATTACHMENT:
+                assert_invariant(false);
+                break;
+        }
+    }
+
+    return buffer.get();
+}
+
 } // namespace backend
 } // namespace filament

filament/backend/src/metal/MetalState.h

@@ -33,32 +33,28 @@
 namespace filament {
 namespace backend {
 
-inline bool operator==(const SamplerParams& lhs, const SamplerParams& rhs) {
-    return SamplerParams::EqualTo{}(lhs, rhs);
-}
-
 //   Rasterization Bindings
 //   ----------------------
 //   Bindings    Buffer name                          Count
 //   ------------------------------------------------------
 //   0           Zero buffer (placeholder vertex buffer)  1
 //   1-16        Filament vertex buffers                 16   limited by MAX_VERTEX_BUFFER_COUNT
-//   17-25       Uniform buffers                          9   Program::UNIFORM_BINDING_COUNT
-//   26          Push constants                           1
-//   27-30       Sampler groups (argument buffers)        4   Program::SAMPLER_BINDING_COUNT
+//   20          Push constants                           1
+//   21-24       Descriptor sets (argument buffers)       4   limited by MAX_DESCRIPTOR_SET_COUNT
+//   25          Dynamic offset buffer                    1
 //
-//   Total                                               31
+//   Total                                               23
 
 //   Compute Bindings
 //   ----------------------
 //   Bindings    Buffer name                          Count
 //   ------------------------------------------------------
 //   0-3         SSBO buffers                             4   MAX_SSBO_COUNT
-//   17-25       Uniform buffers                          9   Program::UNIFORM_BINDING_COUNT
-//   26          Push constants                           1
-//   27-30       Sampler groups (argument buffers)        4   Program::SAMPLER_BINDING_COUNT
+//   20          Push constants                           1
+//   21-24       Descriptor sets (argument buffers)       4   limited by MAX_DESCRIPTOR_SET_COUNT
+//   25          Dynamic offset buffer                    1
 //
-//   Total                                               18
+//   Total                                               10
 
 // The total number of vertex buffer "slots" that the Metal backend can bind.
 // + 1 to account for the zero buffer, a placeholder buffer used internally by the Metal backend.
@@ -71,10 +67,11 @@ static constexpr uint32_t ZERO_VERTEX_BUFFER_BINDING = 0u;
 
 static constexpr uint32_t USER_VERTEX_BUFFER_BINDING_START = 1u;
 
+
 // These constants must match the equivalent in CodeGenerator.h.
-static constexpr uint32_t UNIFORM_BUFFER_BINDING_START = 17u;
-static constexpr uint32_t SSBO_BINDING_START = 0u;
-static constexpr uint32_t SAMPLER_GROUP_BINDING_START = 27u;
+static constexpr uint32_t PUSH_CONSTANT_BUFFER_INDEX = 20u;
+static constexpr uint32_t DESCRIPTOR_SET_BINDING_START = 21u;
+static constexpr uint32_t DYNAMIC_OFFSET_BINDING = 25u;
 
 // Forward declarations necessary here, definitions at end of file.
 inline bool operator==(const MTLViewport& lhs, const MTLViewport& rhs);
@@ -382,18 +379,22 @@ using SamplerStateCache = StateCache<SamplerState, id<MTLSamplerState>, SamplerS
 
 using CullModeStateTracker = StateTracker<MTLCullMode>;
 using WindingStateTracker = StateTracker<MTLWinding>;
+using DepthClampStateTracker = StateTracker<MTLDepthClipMode>;
 
 // Argument encoder
 
 struct ArgumentEncoderState {
+    NSUInteger bufferCount;
     utils::FixedCapacityVector<MTLTextureType> textureTypes;
 
-    explicit ArgumentEncoderState(utils::FixedCapacityVector<MTLTextureType>&& types)
-        : textureTypes(std::move(types)) {}
+    explicit ArgumentEncoderState(
+            NSUInteger bufferCount, utils::FixedCapacityVector<MTLTextureType>&& types)
+        : bufferCount(bufferCount), textureTypes(std::move(types)) {}
 
     bool operator==(const ArgumentEncoderState& rhs) const noexcept {
         return std::equal(textureTypes.begin(), textureTypes.end(), rhs.textureTypes.begin(),
-                rhs.textureTypes.end());
+                       rhs.textureTypes.end()) &&
+                bufferCount == rhs.bufferCount;
     }
 
     bool operator!=(const ArgumentEncoderState& rhs) const noexcept {
@@ -415,6 +416,30 @@ struct ArgumentEncoderCreator {
 using ArgumentEncoderCache = StateCache<ArgumentEncoderState, id<MTLArgumentEncoder>,
         ArgumentEncoderCreator, ArgumentEncoderHasher>;
 
+template <NSUInteger N, ShaderStage stage>
+class MetalBufferBindings {
+public:
+    MetalBufferBindings() { invalidate(); }
+
+    void invalidate() {
+        mDirtyBuffers.reset();
+        mDirtyOffsets.reset();
+        for (int i = 0; i < int(N); i++) {
+            mDirtyBuffers.set(i, true);
+            mDirtyOffsets.set(i, true);
+        }
+    }
+    void setBuffer(const id<MTLBuffer> buffer, NSUInteger offset, NSUInteger index);
+    void bindBuffers(id<MTLCommandEncoder> encoder, NSUInteger startIndex);
+
+private:
+    static_assert(N <= 8);
+    std::array<__weak id<MTLBuffer>, N> mBuffers = { nil };
+    std::array<NSUInteger, N> mOffsets = { 0 };
+    utils::bitset8 mDirtyBuffers;
+    utils::bitset8 mDirtyOffsets;
+};
+
 } // namespace backend
 } // namespace filament
 

filament/backend/src/metal/MetalState.mm

@@ -166,28 +166,40 @@ id<MTLSamplerState> SamplerStateCreator::operator()(id<MTLDevice> device,
 id<MTLArgumentEncoder> ArgumentEncoderCreator::operator()(id<MTLDevice> device,
         const ArgumentEncoderState &state) noexcept {
     const auto& textureTypes = state.textureTypes;
-    const auto& count = textureTypes.size();
-    assert_invariant(count > 0);
+    const auto& textureCount = textureTypes.size();
+    const auto& bufferCount = state.bufferCount;
+    assert_invariant(textureCount > 0);
 
     // Metal has separate data types for textures versus samplers, so the argument buffer layout
     // alternates between texture and sampler, i.e.:
+    // buffer0
+    // buffer1
     // textureA
     // samplerA
     // textureB
     // samplerB
     // etc
     NSMutableArray<MTLArgumentDescriptor*>* arguments =
-            [NSMutableArray arrayWithCapacity:(count * 2)];
-    for (size_t i = 0; i < count; i++) {
+            [NSMutableArray arrayWithCapacity:(bufferCount + textureCount * 2)];
+    size_t i = 0;
+    for (size_t j = 0; j < bufferCount; j++) {
+        MTLArgumentDescriptor* bufferArgument = [MTLArgumentDescriptor argumentDescriptor];
+        bufferArgument.index = i++;
+        bufferArgument.dataType = MTLDataTypePointer;
+        bufferArgument.access = MTLArgumentAccessReadOnly;
+        [arguments addObject:bufferArgument];
+    }
+
+    for (size_t j = 0; j < textureCount; j++) {
         MTLArgumentDescriptor* textureArgument = [MTLArgumentDescriptor argumentDescriptor];
-        textureArgument.index = i * 2 + 0;
+        textureArgument.index = i++;
         textureArgument.dataType = MTLDataTypeTexture;
         textureArgument.textureType = textureTypes[i];
         textureArgument.access = MTLArgumentAccessReadOnly;
         [arguments addObject:textureArgument];
 
         MTLArgumentDescriptor* samplerArgument = [MTLArgumentDescriptor argumentDescriptor];
-        samplerArgument.index = i * 2 + 1;
+        samplerArgument.index = i++;
         samplerArgument.dataType = MTLDataTypeSampler;
         textureArgument.access = MTLArgumentAccessReadOnly;
         [arguments addObject:samplerArgument];
@@ -196,5 +208,64 @@ id<MTLArgumentEncoder> ArgumentEncoderCreator::operator()(id<MTLDevice> device,
     return [device newArgumentEncoderWithArguments:arguments];
 }
 
+template <NSUInteger N, ShaderStage stage>
+void MetalBufferBindings<N, stage>::setBuffer(const id<MTLBuffer> buffer, NSUInteger offset, NSUInteger index) {
+    assert_invariant(offset + 1 <= N);
+
+    if (mBuffers[index] != buffer) {
+        mBuffers[index] = buffer;
+        mDirtyBuffers.set(index);
+    }
+
+    if (mOffsets[index] != offset) {
+        mOffsets[index] = offset;
+        mDirtyOffsets.set(index);
+    }
+}
+
+template <NSUInteger N, ShaderStage stage>
+void MetalBufferBindings<N, stage>::bindBuffers(
+        id<MTLCommandEncoder> encoder, NSUInteger startIndex) {
+    if (mDirtyBuffers.none() && mDirtyOffsets.none()) {
+        return;
+    }
+
+    utils::bitset8 onlyOffsetDirty = mDirtyOffsets & ~mDirtyBuffers;
+    onlyOffsetDirty.forEachSetBit([&](size_t i) {
+        if constexpr (stage == ShaderStage::VERTEX) {
+            [(id<MTLRenderCommandEncoder>)encoder setVertexBufferOffset:mOffsets[i]
+                                                                atIndex:i + startIndex];
+        } else if constexpr (stage == ShaderStage::FRAGMENT) {
+            [(id<MTLRenderCommandEncoder>)encoder setFragmentBufferOffset:mOffsets[i]
+                                                                  atIndex:i + startIndex];
+        } else if constexpr (stage == ShaderStage::COMPUTE) {
+            [(id<MTLComputeCommandEncoder>)encoder setBufferOffset:mOffsets[i]
+                                                           atIndex:i + startIndex];
+        }
+    });
+    mDirtyOffsets.reset();
+
+    mDirtyBuffers.forEachSetBit([&](size_t i) {
+        if constexpr (stage == ShaderStage::VERTEX) {
+            [(id<MTLRenderCommandEncoder>)encoder setVertexBuffer:mBuffers[i]
+                                                           offset:mOffsets[i]
+                                                          atIndex:i + startIndex];
+        } else if constexpr (stage == ShaderStage::FRAGMENT) {
+            [(id<MTLRenderCommandEncoder>)encoder setFragmentBuffer:mBuffers[i]
+                                                             offset:mOffsets[i]
+                                                            atIndex:i + startIndex];
+        } else if constexpr (stage == ShaderStage::COMPUTE) {
+            [(id<MTLComputeCommandEncoder>)encoder setBuffer:mBuffers[i]
+                                                      offset:mOffsets[i]
+                                                     atIndex:i + startIndex];
+        }
+    });
+    mDirtyBuffers.reset();
+}
+
+template class MetalBufferBindings<MAX_DESCRIPTOR_SET_COUNT, ShaderStage::VERTEX>;
+template class MetalBufferBindings<MAX_DESCRIPTOR_SET_COUNT, ShaderStage::FRAGMENT>;
+template class MetalBufferBindings<MAX_DESCRIPTOR_SET_COUNT, ShaderStage::COMPUTE>;
+
 } // namespace backend
 } // namespace filament

filament/backend/src/noop/NoopDriver.cpp

@@ -54,7 +54,7 @@ void NoopDriver::beginFrame(int64_t monotonic_clock_ns,
 }
 
 void NoopDriver::setFrameScheduledCallback(Handle<HwSwapChain> sch,
-        CallbackHandler* handler, FrameScheduledCallback&& callback) {
+        CallbackHandler* handler, FrameScheduledCallback&& callback, uint64_t flags) {
 
 }
 
@@ -99,9 +99,6 @@ void NoopDriver::destroyProgram(Handle<HwProgram> ph) {
 void NoopDriver::destroyRenderTarget(Handle<HwRenderTarget> rth) {
 }
 
-void NoopDriver::destroySamplerGroup(Handle<HwSamplerGroup> sbh) {
-}
-
 void NoopDriver::destroySwapChain(Handle<HwSwapChain> sch) {
 }
 
@@ -111,6 +108,12 @@ void NoopDriver::destroyStream(Handle<HwStream> sh) {
 void NoopDriver::destroyTimerQuery(Handle<HwTimerQuery> tqh) {
 }
 
+void NoopDriver::destroyDescriptorSetLayout(Handle<HwDescriptorSetLayout> tqh) {
+}
+
+void NoopDriver::destroyDescriptorSet(Handle<HwDescriptorSet> tqh) {
+}
+
 Handle<HwStream> NoopDriver::createStreamNative(void* nativeStream) {
     return {};
 }
@@ -202,6 +205,10 @@ bool NoopDriver::isProtectedTexturesSupported() {
     return true;
 }
 
+bool NoopDriver::isDepthClampSupported() {
+    return false;
+}
+
 bool NoopDriver::isWorkaroundNeeded(Workaround) {
     return false;
 }
@@ -244,9 +251,6 @@ void NoopDriver::setVertexBufferObject(Handle<HwVertexBuffer> vbh, uint32_t inde
         Handle<HwBufferObject> boh) {
 }
 
-void NoopDriver::setMinMaxLevels(Handle<HwTexture> th, uint32_t minLevel, uint32_t maxLevel) {
-}
-
 void NoopDriver::update3DImage(Handle<HwTexture> th,
         uint32_t level, uint32_t xoffset, uint32_t yoffset, uint32_t zoffset,
         uint32_t width, uint32_t height, uint32_t depth,
@@ -272,11 +276,6 @@ void NoopDriver::setExternalStream(Handle<HwTexture> th, Handle<HwStream> sh) {
 
 void NoopDriver::generateMipmaps(Handle<HwTexture> th) { }
 
-void NoopDriver::updateSamplerGroup(Handle<HwSamplerGroup> sbh,
-        BufferDescriptor&& data) {
-    scheduleDestroy(std::move(data));
-}
-
 void NoopDriver::compilePrograms(CompilerPriorityQueue priority,
         CallbackHandler* handler, CallbackHandler::Callback callback, void* user) {
     if (callback) {
@@ -299,27 +298,14 @@ void NoopDriver::makeCurrent(Handle<HwSwapChain> drawSch, Handle<HwSwapChain> re
 void NoopDriver::commit(Handle<HwSwapChain> sch) {
 }
 
-void NoopDriver::bindUniformBuffer(uint32_t index, Handle<HwBufferObject> ubh) {
-}
-
-void NoopDriver::bindBufferRange(BufferObjectBinding bindingType, uint32_t index,
-        Handle<HwBufferObject> ubh, uint32_t offset, uint32_t size) {
-}
-
-void NoopDriver::unbindBuffer(BufferObjectBinding bindingType, uint32_t index) {
-}
-
-void NoopDriver::bindSamplers(uint32_t index, Handle<HwSamplerGroup> sbh) {
-}
-
 void NoopDriver::setPushConstant(backend::ShaderStage stage, uint8_t index,
         backend::PushConstantVariant value) {
 }
 
-void NoopDriver::insertEventMarker(char const* string, uint32_t len) {
+void NoopDriver::insertEventMarker(char const* string) {
 }
 
-void NoopDriver::pushGroupMarker(char const* string,  uint32_t len) {
+void NoopDriver::pushGroupMarker(char const* string) {
 }
 
 void NoopDriver::popGroupMarker(int) {
@@ -388,4 +374,28 @@ void NoopDriver::endTimerQuery(Handle<HwTimerQuery> tqh) {
 void NoopDriver::resetState(int) {
 }
 
+void NoopDriver::updateDescriptorSetBuffer(
+        backend::DescriptorSetHandle dsh,
+        backend::descriptor_binding_t binding,
+        backend::BufferObjectHandle boh,
+        uint32_t offset,
+        uint32_t size) {
+}
+
+void NoopDriver::updateDescriptorSetTexture(
+        backend::DescriptorSetHandle dsh,
+        backend::descriptor_binding_t binding,
+        backend::TextureHandle th,
+        SamplerParams params) {
+}
+
+void NoopDriver::bindDescriptorSet(
+        backend::DescriptorSetHandle dsh,
+        backend::descriptor_set_t set,
+        backend::DescriptorSetOffsetArray&& offsets) {
+}
+
+void NoopDriver::setDebugTag(HandleBase::HandleId handleId, utils::CString tag) {
+}
+
 } // namespace filament

filament/backend/src/opengl/BindingMap.h

@@ -0,0 +1,89 @@
+/*
+ * Copyright (C) 2024 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef TNT_FILAMENT_BACKEND_OPENGL_BINDINGMAP_H
+#define TNT_FILAMENT_BACKEND_OPENGL_BINDINGMAP_H
+
+#include <backend/DriverEnums.h>
+
+#include "gl_headers.h"
+
+#include <utils/bitset.h>
+#include <utils/debug.h>
+
+#include <new>
+
+#include <stddef.h>
+#include <stdint.h>
+#include <string.h>
+
+namespace filament::backend {
+
+class BindingMap {
+    struct CompressedBinding {
+        // this is in fact a GLuint, but we only want 8-bits
+        uint8_t binding : 7;
+        uint8_t sampler : 1;
+    };
+
+    CompressedBinding (*mStorage)[MAX_DESCRIPTOR_COUNT];
+
+    utils::bitset64 mActiveDescriptors[MAX_DESCRIPTOR_SET_COUNT];
+
+public:
+    BindingMap() noexcept
+            : mStorage(new (std::nothrow) CompressedBinding[MAX_DESCRIPTOR_SET_COUNT][MAX_DESCRIPTOR_COUNT]) {
+#ifndef NDEBUG
+        memset(mStorage, 0xFF, sizeof(CompressedBinding[MAX_DESCRIPTOR_SET_COUNT][MAX_DESCRIPTOR_COUNT]));
+#endif
+    }
+
+    ~BindingMap() noexcept {
+        delete [] mStorage;
+    }
+
+    BindingMap(BindingMap const&) noexcept = delete;
+    BindingMap(BindingMap&&) noexcept = delete;
+    BindingMap& operator=(BindingMap const&) noexcept = delete;
+    BindingMap& operator=(BindingMap&&) noexcept = delete;
+
+    struct Binding {
+        GLuint binding;
+        DescriptorType type;
+    };
+
+    void insert(descriptor_set_t set, descriptor_binding_t binding, Binding entry) noexcept {
+        assert_invariant(set < MAX_DESCRIPTOR_SET_COUNT);
+        assert_invariant(binding < MAX_DESCRIPTOR_COUNT);
+        assert_invariant(entry.binding < 128); // we reserve 1 bit for the type right now
+        mStorage[set][binding] = { (uint8_t)entry.binding, entry.type == DescriptorType::SAMPLER };
+        mActiveDescriptors[set].set(binding);
+    }
+
+    GLuint get(descriptor_set_t set, descriptor_binding_t binding) const noexcept {
+        assert_invariant(set < MAX_DESCRIPTOR_SET_COUNT);
+        assert_invariant(binding < MAX_DESCRIPTOR_COUNT);
+        return mStorage[set][binding].binding;
+    }
+
+    utils::bitset64 getActiveDescriptors(descriptor_set_t set) const noexcept {
+        return mActiveDescriptors[set];
+    }
+};
+
+} // namespace filament::backend
+
+#endif //TNT_FILAMENT_BACKEND_OPENGL_BINDINGMAP_H

filament/backend/src/opengl/GLDescriptorSet.cpp

@@ -0,0 +1,363 @@
+/*
+ * Copyright (C) 2024 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "GLDescriptorSet.h"
+
+#include "GLBufferObject.h"
+#include "GLDescriptorSetLayout.h"
+#include "GLTexture.h"
+#include "GLUtils.h"
+#include "OpenGLDriver.h"
+#include "OpenGLContext.h"
+#include "OpenGLProgram.h"
+
+#include "gl_headers.h"
+
+#include <private/backend/HandleAllocator.h>
+
+#include <backend/DriverEnums.h>
+#include <backend/Handle.h>
+
+#include <utils/BitmaskEnum.h>
+#include <utils/Log.h>
+#include <utils/Panic.h>
+#include <utils/bitset.h>
+#include <utils/compiler.h>
+#include <utils/debug.h>
+#include <algorithm>
+
+#include <type_traits>
+#include <utility>
+#include <variant>
+
+#include <stddef.h>
+#include <stdint.h>
+
+namespace filament::backend {
+
+GLDescriptorSet::GLDescriptorSet(OpenGLContext& gl, DescriptorSetLayoutHandle dslh,
+        GLDescriptorSetLayout const* layout) noexcept
+        : descriptors(layout->maxDescriptorBinding + 1),
+          dslh(std::move(dslh)) {
+
+    // We have allocated enough storage for all descriptors. Now allocate the empty descriptor
+    // themselves.
+    for (auto const& entry : layout->bindings) {
+        size_t const index = entry.binding;
+
+        // now we'll initialize the alternative for each way we can handle this descriptor.
+        auto& desc = descriptors[index].desc;
+        switch (entry.type) {
+            case DescriptorType::UNIFORM_BUFFER: {
+                // A uniform buffer can have dynamic offsets or not and have special handling for
+                // ES2 (where we need to emulate it). That's four alternatives.
+                bool const dynamicOffset = any(entry.flags & DescriptorFlags::DYNAMIC_OFFSET);
+                dynamicBuffers.set(index, dynamicOffset);
+                if (UTILS_UNLIKELY(gl.isES2())) {
+                    if (dynamicOffset) {
+                        dynamicBufferCount++;
+                    }
+                    desc.emplace<BufferGLES2>(dynamicOffset);
+                } else {
+                    auto const type = GLUtils::getBufferBindingType(BufferObjectBinding::UNIFORM);
+                    if (dynamicOffset) {
+                        dynamicBufferCount++;
+                        desc.emplace<DynamicBuffer>(type);
+                    } else {
+                        desc.emplace<Buffer>(type);
+                    }
+                }
+                break;
+            }
+            case DescriptorType::SHADER_STORAGE_BUFFER: {
+                // shader storage buffers are not supported on ES2, So that's two alternatives.
+                bool const dynamicOffset = any(entry.flags & DescriptorFlags::DYNAMIC_OFFSET);
+                dynamicBuffers.set(index, dynamicOffset);
+                auto const type = GLUtils::getBufferBindingType(BufferObjectBinding::SHADER_STORAGE);
+                if (dynamicOffset) {
+                    dynamicBufferCount++;
+                    desc.emplace<DynamicBuffer>(type);
+                } else {
+                    desc.emplace<Buffer>(type);
+                }
+                break;
+            }
+            case DescriptorType::SAMPLER:
+                if (UTILS_UNLIKELY(gl.isES2())) {
+                    desc.emplace<SamplerGLES2>();
+                } else {
+                    const bool anisotropyWorkaround =
+                            gl.ext.EXT_texture_filter_anisotropic &&
+                            gl.bugs.texture_filter_anisotropic_broken_on_sampler;
+                    if (anisotropyWorkaround) {
+                        desc.emplace<SamplerWithAnisotropyWorkaround>();
+                    } else {
+                        desc.emplace<Sampler>();
+                    }
+                }
+                break;
+            case DescriptorType::INPUT_ATTACHMENT:
+                break;
+        }
+    }
+}
+
+void GLDescriptorSet::update(OpenGLContext&,
+        descriptor_binding_t binding, GLBufferObject* bo, size_t offset, size_t size) noexcept {
+    assert_invariant(binding < descriptors.size());
+    std::visit([=](auto&& arg) {
+        using T = std::decay_t<decltype(arg)>;
+        if constexpr (std::is_same_v<T, Buffer> || std::is_same_v<T, DynamicBuffer>) {
+            assert_invariant(arg.target != 0);
+            arg.id = bo ? bo->gl.id : 0;
+            arg.offset = uint32_t(offset);
+            arg.size = uint32_t(size);
+            assert_invariant(arg.id || (!arg.size && !offset));
+        } else if constexpr (std::is_same_v<T, BufferGLES2>) {
+            arg.bo = bo;
+            arg.offset = uint32_t(offset);
+        } else {
+            // API usage error. User asked to update the wrong type of descriptor.
+            PANIC_PRECONDITION("descriptor %d is not a buffer", +binding);
+        }
+    }, descriptors[binding].desc);
+}
+
+void GLDescriptorSet::update(OpenGLContext& gl,
+        descriptor_binding_t binding, GLTexture* t, SamplerParams params) noexcept {
+    assert_invariant(binding < descriptors.size());
+    std::visit([=, &gl](auto&& arg) mutable {
+        using T = std::decay_t<decltype(arg)>;
+        if constexpr (std::is_same_v<T, Sampler> ||
+                      std::is_same_v<T, SamplerWithAnisotropyWorkaround> ||
+                      std::is_same_v<T, SamplerGLES2>) {
+            if (UTILS_UNLIKELY(t && t->target == SamplerType::SAMPLER_EXTERNAL)) {
+                // From OES_EGL_image_external spec:
+                // "The default s and t wrap modes are CLAMP_TO_EDGE, and it is an INVALID_ENUM
+                //  error to set the wrap mode to any other value."
+                params.wrapS = SamplerWrapMode::CLAMP_TO_EDGE;
+                params.wrapT = SamplerWrapMode::CLAMP_TO_EDGE;
+                params.wrapR = SamplerWrapMode::CLAMP_TO_EDGE;
+            }
+            // GLES3.x specification forbids depth textures to be filtered.
+            if (t && isDepthFormat(t->format)
+                    && params.compareMode == SamplerCompareMode::NONE) {
+                params.filterMag = SamplerMagFilter::NEAREST;
+                switch (params.filterMin) {
+                    case SamplerMinFilter::LINEAR:
+                        params.filterMin = SamplerMinFilter::NEAREST;
+                        break;
+                    case SamplerMinFilter::LINEAR_MIPMAP_NEAREST:
+                    case SamplerMinFilter::NEAREST_MIPMAP_LINEAR:
+                    case SamplerMinFilter::LINEAR_MIPMAP_LINEAR:
+                        params.filterMin = SamplerMinFilter::NEAREST_MIPMAP_NEAREST;
+                        break;
+                    default:
+                        break;
+                }
+            }
+
+            arg.target = t ? t->gl.target : 0;
+            arg.id = t ? t->gl.id : 0;
+            if constexpr (std::is_same_v<T, Sampler> ||
+                          std::is_same_v<T, SamplerWithAnisotropyWorkaround>) {
+                if constexpr (std::is_same_v<T, SamplerWithAnisotropyWorkaround>) {
+                    arg.anisotropy = float(1u << params.anisotropyLog2);
+                }
+                if (t) {
+                    arg.ref = t->ref;
+                    arg.baseLevel = t->gl.baseLevel;
+                    arg.maxLevel = t->gl.maxLevel;
+                    arg.swizzle = t->gl.swizzle;
+                }
+#ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
+                arg.sampler = gl.getSampler(params);
+#else
+                (void)gl;
+#endif
+            } else {
+                arg.params = params;
+            }
+        } else {
+            // API usage error. User asked to update the wrong type of descriptor.
+            PANIC_PRECONDITION("descriptor %d is not a texture", +binding);
+        }
+    }, descriptors[binding].desc);
+}
+
+template<typename T>
+void GLDescriptorSet::updateTextureView(OpenGLContext& gl,
+        HandleAllocatorGL& handleAllocator, GLuint unit, T const& desc) noexcept {
+    // The common case is that we don't have a ref handle (we only have one if
+    // the texture ever had a View on it).
+    assert_invariant(desc.ref);
+    GLTextureRef* const ref = handleAllocator.handle_cast<GLTextureRef*>(desc.ref);
+    if (UTILS_UNLIKELY((desc.baseLevel != ref->baseLevel || desc.maxLevel != ref->maxLevel))) {
+        // If we have views, then it's still uncommon that we'll switch often
+        // handle the case where we reset to the original texture
+        GLint baseLevel = GLint(desc.baseLevel); // NOLINT(*-signed-char-misuse)
+        GLint maxLevel = GLint(desc.maxLevel); // NOLINT(*-signed-char-misuse)
+        if (baseLevel > maxLevel) {
+            baseLevel = 0;
+            maxLevel = 1000; // per OpenGL spec
+        }
+        // that is very unfortunate that we have to call activeTexture here
+        gl.activeTexture(unit);
+        glTexParameteri(desc.target, GL_TEXTURE_BASE_LEVEL, baseLevel);
+        glTexParameteri(desc.target, GL_TEXTURE_MAX_LEVEL,  maxLevel);
+        ref->baseLevel = desc.baseLevel;
+        ref->maxLevel = desc.maxLevel;
+    }
+    if (UTILS_UNLIKELY(desc.swizzle != ref->swizzle)) {
+        using namespace GLUtils;
+        gl.activeTexture(unit);
+#if !defined(__EMSCRIPTEN__)  && !defined(FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2)
+        glTexParameteri(desc.target, GL_TEXTURE_SWIZZLE_R, (GLint)getSwizzleChannel(desc.swizzle[0]));
+        glTexParameteri(desc.target, GL_TEXTURE_SWIZZLE_G, (GLint)getSwizzleChannel(desc.swizzle[1]));
+        glTexParameteri(desc.target, GL_TEXTURE_SWIZZLE_B, (GLint)getSwizzleChannel(desc.swizzle[2]));
+        glTexParameteri(desc.target, GL_TEXTURE_SWIZZLE_A, (GLint)getSwizzleChannel(desc.swizzle[3]));
+#endif
+        ref->swizzle = desc.swizzle;
+    }
+}
+
+void GLDescriptorSet::bind(
+        OpenGLContext& gl,
+        HandleAllocatorGL& handleAllocator,
+        OpenGLProgram const& p,
+        descriptor_set_t set, uint32_t const* offsets, bool offsetsOnly) const noexcept {
+    // TODO: check that offsets is sized correctly
+    size_t dynamicOffsetIndex = 0;
+
+    utils::bitset64 activeDescriptorBindings = p.getActiveDescriptors(set);
+    if (offsetsOnly) {
+        activeDescriptorBindings &= dynamicBuffers;
+    }
+
+    // loop only over the active indices for this program
+    activeDescriptorBindings.forEachSetBit(
+            [this,&gl, &handleAllocator, &p, set, offsets, &dynamicOffsetIndex]
+            (size_t binding) {
+
+        // This would fail here if we're trying to set a descriptor that doesn't exist in the
+        // program. In other words, a mismatch between the program's layout and this descriptor-set.
+        assert_invariant(binding < descriptors.size());
+
+        auto const& entry = descriptors[binding];
+        std::visit(
+                [&gl, &handleAllocator, &p, &dynamicOffsetIndex, set, binding, offsets]
+                (auto&& arg) {
+            using T = std::decay_t<decltype(arg)>;
+            if constexpr (std::is_same_v<T, Buffer>) {
+                GLuint const bindingPoint = p.getBufferBinding(set, binding);
+                GLintptr const offset = arg.offset;
+                assert_invariant(arg.id || (!arg.size && !offset));
+                gl.bindBufferRange(arg.target, bindingPoint, arg.id, offset, arg.size);
+            } else if constexpr (std::is_same_v<T, DynamicBuffer>) {
+                GLuint const bindingPoint = p.getBufferBinding(set, binding);
+                GLintptr const offset = arg.offset + offsets[dynamicOffsetIndex++];
+                assert_invariant(arg.id || (!arg.size && !offset));
+                gl.bindBufferRange(arg.target, bindingPoint, arg.id, offset, arg.size);
+            } else if constexpr (std::is_same_v<T, BufferGLES2>) {
+                GLuint const bindingPoint = p.getBufferBinding(set, binding);
+                GLintptr offset = arg.offset;
+                if (arg.dynamicOffset) {
+                    offset += offsets[dynamicOffsetIndex++];
+                }
+                if (arg.bo) {
+                    auto buffer = static_cast<char const*>(arg.bo->gl.buffer) + offset;
+                    p.updateUniforms(bindingPoint, arg.bo->gl.id, buffer, arg.bo->age);
+                }
+            } else if constexpr (std::is_same_v<T, Sampler>) {
+                GLuint const unit = p.getTextureUnit(set, binding);
+                if (arg.target) {
+                    gl.bindTexture(unit, arg.target, arg.id);
+                    gl.bindSampler(unit, arg.sampler);
+                    if (UTILS_UNLIKELY(arg.ref)) {
+                        updateTextureView(gl, handleAllocator, unit, arg);
+                    }
+                } else {
+                    gl.unbindTextureUnit(unit);
+                }
+            } else if constexpr (std::is_same_v<T, SamplerWithAnisotropyWorkaround>) {
+                GLuint const unit = p.getTextureUnit(set, binding);
+                if (arg.target) {
+                    gl.bindTexture(unit, arg.target, arg.id);
+                    gl.bindSampler(unit, arg.sampler);
+                    if (UTILS_UNLIKELY(arg.ref)) {
+                        updateTextureView(gl, handleAllocator, unit, arg);
+                    }
+#if defined(GL_EXT_texture_filter_anisotropic)
+                    // Driver claims to support anisotropic filtering, but it fails when set on
+                    // the sampler, we have to set it on the texture instead.
+                    glTexParameterf(arg.target, GL_TEXTURE_MAX_ANISOTROPY_EXT,
+                            std::min(gl.gets.max_anisotropy, float(arg.anisotropy)));
+#endif
+                } else {
+                    gl.unbindTextureUnit(unit);
+                }
+            } else if constexpr (std::is_same_v<T, SamplerGLES2>) {
+                // in ES2 the sampler parameters need to be set on the texture itself
+                GLuint const unit = p.getTextureUnit(set, binding);
+                if (arg.target) {
+                    gl.bindTexture(unit, arg.target, arg.id);
+                    SamplerParams const params = arg.params;
+                    glTexParameteri(arg.target, GL_TEXTURE_MIN_FILTER,
+                            (GLint)GLUtils::getTextureFilter(params.filterMin));
+                    glTexParameteri(arg.target, GL_TEXTURE_MAG_FILTER,
+                            (GLint)GLUtils::getTextureFilter(params.filterMag));
+                    glTexParameteri(arg.target, GL_TEXTURE_WRAP_S,
+                            (GLint)GLUtils::getWrapMode(params.wrapS));
+                    glTexParameteri(arg.target, GL_TEXTURE_WRAP_T,
+                            (GLint)GLUtils::getWrapMode(params.wrapT));
+#if defined(GL_EXT_texture_filter_anisotropic)
+                    glTexParameterf(arg.target, GL_TEXTURE_MAX_ANISOTROPY_EXT,
+                            std::min(gl.gets.max_anisotropy, arg.anisotropy));
+#endif
+                } else {
+                    gl.unbindTextureUnit(unit);
+                }
+            }
+        }, entry.desc);
+    });
+    CHECK_GL_ERROR(utils::slog.e)
+}
+
+void GLDescriptorSet::validate(HandleAllocatorGL& allocator,
+        DescriptorSetLayoutHandle pipelineLayout) const {
+
+    if (UTILS_UNLIKELY(dslh != pipelineLayout)) {
+        auto* const dsl = allocator.handle_cast < GLDescriptorSetLayout const * > (dslh);
+        auto* const cur = allocator.handle_cast < GLDescriptorSetLayout const * > (pipelineLayout);
+
+        UTILS_UNUSED_IN_RELEASE
+        bool const pipelineLayoutMatchesDescriptorSetLayout = std::equal(
+                dsl->bindings.begin(), dsl->bindings.end(),
+                cur->bindings.begin(),
+                [](DescriptorSetLayoutBinding const& lhs,
+                        DescriptorSetLayoutBinding const& rhs) {
+                    return lhs.type == rhs.type &&
+                           lhs.stageFlags == rhs.stageFlags &&
+                           lhs.binding == rhs.binding &&
+                           lhs.flags == rhs.flags &&
+                           lhs.count == rhs.count;
+                });
+
+        assert_invariant(pipelineLayoutMatchesDescriptorSetLayout);
+    }
+}
+
+} // namespace filament::backend

filament/backend/src/opengl/GLDescriptorSet.h

@@ -0,0 +1,175 @@
+/*
+ * Copyright (C) 2024 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef TNT_FILAMENT_BACKEND_OPENGL_GLDESCRIPTORSET_H
+#define TNT_FILAMENT_BACKEND_OPENGL_GLDESCRIPTORSET_H
+
+#include "DriverBase.h"
+
+#include "gl_headers.h"
+
+#include <private/backend/HandleAllocator.h>
+
+#include <backend/DriverEnums.h>
+#include <backend/Handle.h>
+
+#include <utils/bitset.h>
+#include <utils/FixedCapacityVector.h>
+
+#include <math/half.h>
+
+#include <array>
+#include <variant>
+
+#include <stddef.h>
+#include <stdint.h>
+
+namespace filament::backend {
+
+struct GLBufferObject;
+struct GLTexture;
+struct GLTextureRef;
+struct GLDescriptorSetLayout;
+class OpenGLProgram;
+class OpenGLContext;
+class OpenGLDriver;
+
+struct GLDescriptorSet : public HwDescriptorSet {
+
+    using HwDescriptorSet::HwDescriptorSet;
+
+    GLDescriptorSet(OpenGLContext& gl, DescriptorSetLayoutHandle dslh,
+            GLDescriptorSetLayout const* layout) noexcept;
+
+    // update a buffer descriptor in the set
+    void update(OpenGLContext& gl,
+            descriptor_binding_t binding, GLBufferObject* bo, size_t offset, size_t size) noexcept;
+
+    // update a sampler descriptor in the set
+    void update(OpenGLContext& gl,
+            descriptor_binding_t binding, GLTexture* t, SamplerParams params) noexcept;
+
+    // conceptually bind the set to the command buffer
+    void bind(
+            OpenGLContext& gl,
+            HandleAllocatorGL& handleAllocator,
+            OpenGLProgram const& p,
+            descriptor_set_t set, uint32_t const* offsets, bool offsetsOnly) const noexcept;
+
+    uint32_t getDynamicBufferCount() const noexcept {
+        return dynamicBufferCount;
+    }
+
+    void validate(HandleAllocatorGL& allocator, DescriptorSetLayoutHandle pipelineLayout) const;
+
+private:
+    // a Buffer Descriptor such as SSBO or UBO with static offset
+    struct Buffer {
+        // Workaround: we cannot define the following as Buffer() = default because one of our
+        // clients has their compiler set up where such declaration (possibly coupled with explicit)
+        // will be considered a deleted constructor.
+        Buffer() {}
+
+        explicit Buffer(GLenum target) noexcept : target(target) {}
+        GLenum target;                          // 4
+        GLuint id = 0;                          // 4
+        uint32_t offset = 0;                    // 4
+        uint32_t size = 0;                      // 4
+    };
+
+    // a Buffer Descriptor such as SSBO or UBO with dynamic offset
+    struct DynamicBuffer {
+        DynamicBuffer() = default;
+        explicit DynamicBuffer(GLenum target) noexcept : target(target) { }
+        GLenum target;                          // 4
+        GLuint id = 0;                          // 4
+        uint32_t offset = 0;                    // 4
+        uint32_t size = 0;                      // 4
+    };
+
+    // a UBO descriptor for ES2
+    struct BufferGLES2 {
+        BufferGLES2() = default;
+        explicit BufferGLES2(bool dynamicOffset) noexcept : dynamicOffset(dynamicOffset) { }
+        GLBufferObject const* bo = nullptr;     // 8
+        uint32_t offset = 0;                    // 4
+        bool dynamicOffset = false;             // 4
+    };
+
+    // A sampler descriptor
+    struct Sampler {
+        GLenum target = 0;                      // 4
+        GLuint id = 0;                          // 4
+        GLuint sampler = 0;                     // 4
+        Handle<GLTextureRef> ref;               // 4
+        int8_t baseLevel = 0x7f;                // 1
+        int8_t maxLevel = -1;                   // 1
+        std::array<TextureSwizzle, 4> swizzle{  // 4
+                TextureSwizzle::CHANNEL_0,
+                TextureSwizzle::CHANNEL_1,
+                TextureSwizzle::CHANNEL_2,
+                TextureSwizzle::CHANNEL_3
+        };
+    };
+
+    struct SamplerWithAnisotropyWorkaround {
+        GLenum target = 0;                      // 4
+        GLuint id = 0;                          // 4
+        GLuint sampler = 0;                     // 4
+        Handle<GLTextureRef> ref;               // 4
+        math::half anisotropy = 1.0f;           // 2
+        int8_t baseLevel = 0x7f;                // 1
+        int8_t maxLevel = -1;                   // 1
+        std::array<TextureSwizzle, 4> swizzle{  // 4
+                TextureSwizzle::CHANNEL_0,
+                TextureSwizzle::CHANNEL_1,
+                TextureSwizzle::CHANNEL_2,
+                TextureSwizzle::CHANNEL_3
+        };
+    };
+
+    // A sampler descriptor for ES2
+    struct SamplerGLES2 {
+        GLenum target = 0;                      // 4
+        GLuint id = 0;                          // 4
+        SamplerParams params{};                 // 4
+        float anisotropy = 1.0f;                // 4
+    };
+    struct Descriptor {
+        std::variant<
+                Buffer,
+                DynamicBuffer,
+                BufferGLES2,
+                Sampler,
+                SamplerWithAnisotropyWorkaround,
+                SamplerGLES2> desc;
+    };
+    static_assert(sizeof(Descriptor) <= 32);
+
+    template<typename T>
+    static void updateTextureView(OpenGLContext& gl,
+            HandleAllocatorGL& handleAllocator, GLuint unit, T const& desc) noexcept;
+
+    utils::FixedCapacityVector<Descriptor> descriptors;     // 16
+    utils::bitset64 dynamicBuffers;                         // 8
+    DescriptorSetLayoutHandle dslh;                         // 4
+    uint8_t dynamicBufferCount = 0;                         // 1
+};
+static_assert(sizeof(GLDescriptorSet) <= 32);
+
+} // namespace filament::backend
+
+#endif //TNT_FILAMENT_BACKEND_OPENGL_GLDESCRIPTORSET_H

filament/backend/src/opengl/GLDescriptorSetLayout.h

@@ -0,0 +1,52 @@
+/*
+ * Copyright (C) 2024 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef TNT_FILAMENT_BACKEND_OPENGL_GLDESCRIPTORSETLAYOUT_H
+#define TNT_FILAMENT_BACKEND_OPENGL_GLDESCRIPTORSETLAYOUT_H
+
+#include "DriverBase.h"
+
+#include <backend/DriverEnums.h>
+
+#include <algorithm>
+#include <utility>
+
+#include <stdint.h>
+
+namespace filament::backend {
+
+struct GLDescriptorSetLayout : public HwDescriptorSetLayout, public DescriptorSetLayout {
+    using HwDescriptorSetLayout::HwDescriptorSetLayout;
+    explicit GLDescriptorSetLayout(DescriptorSetLayout&& layout) noexcept
+            : DescriptorSetLayout(std::move(layout)) {
+
+        std::sort(bindings.begin(), bindings.end(),
+                [](auto&& lhs, auto&& rhs){
+            return lhs.binding < rhs.binding;
+        });
+
+        auto p = std::max_element(bindings.cbegin(), bindings.cend(),
+                [](auto const& lhs, auto const& rhs) {
+            return lhs.binding < rhs.binding;
+        });
+        maxDescriptorBinding = p->binding;
+    }
+    uint8_t maxDescriptorBinding = 0;
+};
+
+} // namespace filament::backend
+
+#endif //TNT_FILAMENT_BACKEND_OPENGL_GLDESCRIPTORSETLAYOUT_H

filament/backend/src/opengl/GLTexture.h

@@ -21,12 +21,32 @@
 
 #include "gl_headers.h"
 
+#include <backend/Handle.h>
+#include <backend/DriverEnums.h>
 #include <backend/platforms/OpenGLPlatform.h>
 
+#include <array>
+
 #include <stdint.h>
 
 namespace filament::backend {
 
+struct GLTextureRef {
+    GLTextureRef() = default;
+    // view reference counter
+    uint16_t count = 1;
+    // current per-view values of the texture (in GL we can only have a single View active at
+    // a time, and this tracks that state). It's used to avoid unnecessarily change state.
+    int8_t baseLevel = 127;
+    int8_t maxLevel = -1;
+    std::array<TextureSwizzle, 4> swizzle{
+            TextureSwizzle::CHANNEL_0,
+            TextureSwizzle::CHANNEL_1,
+            TextureSwizzle::CHANNEL_2,
+            TextureSwizzle::CHANNEL_3
+    };
+};
+
 struct GLTexture : public HwTexture {
     using HwTexture::HwTexture;
     struct GL {
@@ -44,8 +64,14 @@ struct GLTexture : public HwTexture {
         bool imported           : 1;
         uint8_t sidecarSamples  : 4;
         uint8_t reserved1       : 3;
+        std::array<TextureSwizzle, 4> swizzle{
+                TextureSwizzle::CHANNEL_0,
+                TextureSwizzle::CHANNEL_1,
+                TextureSwizzle::CHANNEL_2,
+                TextureSwizzle::CHANNEL_3
+        };
     } gl;
-
+    mutable Handle<GLTextureRef> ref;
     OpenGLPlatform::ExternalTexture* externalTexture = nullptr;
 };
 

filament/backend/src/opengl/OpenGLContext.cpp

@@ -552,6 +552,14 @@ void OpenGLContext::initBugs(Bugs* bugs, Extensions const& exts,
         } else if (strstr(renderer, "AMD") ||
                    strstr(renderer, "ATI")) {
             // AMD/ATI GPU
+        } else if (strstr(vendor, "Mesa")) {
+            // Seen on
+            //  [Mesa],
+            //  [llvmpipe (LLVM 17.0.6, 256 bits)],
+            //  [4.5 (Core Profile) Mesa 24.0.6-1],
+            //  [4.50]
+            // not known which version are affected
+            bugs->rebind_buffer_after_deletion = true;
         } else if (strstr(renderer, "Mozilla")) {
             bugs->disable_invalidate_framebuffer = true;
         }
@@ -676,6 +684,7 @@ void OpenGLContext::initExtensionsGLES(Extensions* ext, GLint major, GLint minor
 #ifndef __EMSCRIPTEN__
     ext->EXT_debug_marker = exts.has("GL_EXT_debug_marker"sv);
 #endif
+    ext->EXT_depth_clamp = exts.has("GL_EXT_depth_clamp"sv);
     ext->EXT_discard_framebuffer = exts.has("GL_EXT_discard_framebuffer"sv);
 #ifndef __EMSCRIPTEN__
     ext->EXT_disjoint_timer_query = exts.has("GL_EXT_disjoint_timer_query"sv);
@@ -746,6 +755,7 @@ void OpenGLContext::initExtensionsGL(Extensions* ext, GLint major, GLint minor)
     ext->EXT_color_buffer_half_float = true;  // Assumes core profile.
     ext->EXT_clip_cull_distance = true;
     ext->EXT_debug_marker = exts.has("GL_EXT_debug_marker"sv);
+    ext->EXT_depth_clamp = true;
     ext->EXT_discard_framebuffer = false;
     ext->EXT_disjoint_timer_query = true;
     ext->EXT_multisampled_render_to_texture = false;
@@ -927,15 +937,19 @@ void OpenGLContext::unbindSampler(GLuint sampler) noexcept {
     }
 }
 
-void OpenGLContext::deleteBuffers(GLsizei n, const GLuint* buffers, GLenum target) noexcept {
-    glDeleteBuffers(n, buffers);
+void OpenGLContext::deleteBuffer(GLuint buffer, GLenum target) noexcept {
+    glDeleteBuffers(1, &buffer);
+
     // bindings of bound buffers are reset to 0
-    const size_t targetIndex = getIndexForBufferTarget(target);
-    auto& genericBuffer = state.buffers.genericBinding[targetIndex];
-    UTILS_NOUNROLL
-    for (GLsizei i = 0; i < n; ++i) {
-        if (genericBuffer == buffers[i]) {
-            genericBuffer = 0;
+    size_t const targetIndex = getIndexForBufferTarget(target);
+    auto& genericBinding = state.buffers.genericBinding[targetIndex];
+    if (genericBinding == buffer) {
+        genericBinding = 0;
+    }
+
+    if (UTILS_UNLIKELY(bugs.rebind_buffer_after_deletion)) {
+        if (genericBinding) {
+            glBindBuffer(target, genericBinding);
         }
     }
 
@@ -944,16 +958,13 @@ void OpenGLContext::deleteBuffers(GLsizei n, const GLuint* buffers, GLenum targe
             (target != GL_UNIFORM_BUFFER && target != GL_TRANSFORM_FEEDBACK_BUFFER));
 
     if (target == GL_UNIFORM_BUFFER || target == GL_TRANSFORM_FEEDBACK_BUFFER) {
-        auto& indexedBuffer = state.buffers.targets[targetIndex];
-        UTILS_NOUNROLL // clang generates >1 KiB of code!!
-        for (GLsizei i = 0; i < n; ++i) {
-            UTILS_NOUNROLL
-            for (auto& buffer : indexedBuffer.buffers) {
-                if (buffer.name == buffers[i]) {
-                    buffer.name = 0;
-                    buffer.offset = 0;
-                    buffer.size = 0;
-                }
+        auto& indexedBinding = state.buffers.targets[targetIndex];
+        UTILS_NOUNROLL
+        for (auto& entry: indexedBinding.buffers) {
+            if (entry.name == buffer) {
+                entry.name = 0;
+                entry.offset = 0;
+                entry.size = 0;
             }
         }
     }

filament/backend/src/opengl/OpenGLContext.h

@@ -60,10 +60,19 @@ public:
     struct RenderPrimitive {
         static_assert(MAX_VERTEX_ATTRIBUTE_COUNT <= 16);
 
-        GLuint vao[2] = {};                                     // 4
+        GLuint vao[2] = {};                                     // 8
         GLuint elementArray = 0;                                // 4
+        GLenum indicesType = 0;                                 // 4
+
+        // The optional 32-bit handle to a GLVertexBuffer is necessary only if the referenced
+        // VertexBuffer supports buffer objects. If this is zero, then the VBO handles array is
+        // immutable.
+        Handle<HwVertexBuffer> vertexBufferWithObjects;         // 4
+
         mutable utils::bitset<uint16_t> vertexAttribArray;      // 2
 
+        uint8_t reserved[2] = {};                               // 2
+
         // if this differs from vertexBufferWithObjects->bufferObjectsVersion, this VAO needs to
         // be updated (see OpenGLDriver::updateVertexArrayObject())
         uint8_t vertexBufferVersion = 0;                        // 1
@@ -76,16 +85,11 @@ public:
         // See OpenGLContext::bindVertexArray()
         uint8_t nameVersion = 0;                                // 1
 
-        // Size in bytes of indices in the index buffer
-        uint8_t indicesSize = 0;                                // 1
-
-        // The optional 32-bit handle to a GLVertexBuffer is necessary only if the referenced
-        // VertexBuffer supports buffer objects. If this is zero, then the VBO handles array is
-        // immutable.
-        Handle<HwVertexBuffer> vertexBufferWithObjects;         // 4
+        // Size in bytes of indices in the index buffer (1 or 2)
+        uint8_t indicesShift = 0;                                // 1
 
         GLenum getIndicesType() const noexcept {
-            return indicesSize == 4 ? GL_UNSIGNED_INT : GL_UNSIGNED_SHORT;
+            return indicesType;
         }
     } gl;
 
@@ -186,7 +190,7 @@ public:
     inline void viewport(GLint left, GLint bottom, GLsizei width, GLsizei height) noexcept;
     inline void depthRange(GLclampf near, GLclampf far) noexcept;
 
-    void deleteBuffers(GLsizei n, const GLuint* buffers, GLenum target) noexcept;
+    void deleteBuffer(GLuint buffer, GLenum target) noexcept;
     void deleteVertexArray(GLuint vao) noexcept;
 
     void destroyWithContext(size_t index, std::function<void(OpenGLContext&)> const& closure) noexcept;
@@ -220,8 +224,9 @@ public:
         bool EXT_color_buffer_float;
         bool EXT_color_buffer_half_float;
         bool EXT_debug_marker;
-        bool EXT_disjoint_timer_query;
+        bool EXT_depth_clamp;
         bool EXT_discard_framebuffer;
+        bool EXT_disjoint_timer_query;
         bool EXT_multisampled_render_to_texture2;
         bool EXT_multisampled_render_to_texture;
         bool EXT_protected_textures;
@@ -239,10 +244,10 @@ public:
         bool KHR_parallel_shader_compile;
         bool KHR_texture_compression_astc_hdr;
         bool KHR_texture_compression_astc_ldr;
-        bool OES_depth_texture;
-        bool OES_depth24;
-        bool OES_packed_depth_stencil;
         bool OES_EGL_image_external_essl3;
+        bool OES_depth24;
+        bool OES_depth_texture;
+        bool OES_packed_depth_stencil;
         bool OES_rgb8_rgba8;
         bool OES_standard_derivatives;
         bool OES_texture_npot;
@@ -311,10 +316,15 @@ public:
         // a glFinish. So we must delay the destruction until we know the GPU is finished.
         bool delay_fbo_destruction;
 
+        // Mesa sometimes clears the generic buffer binding when *another* buffer is destroyed,
+        // if that other buffer is bound on an *indexed* buffer binding.
+        bool rebind_buffer_after_deletion;
+
         // Force feature level 0. Typically used for low end ES3 devices with significant driver
         // bugs or performance issues.
         bool force_feature_level0;
 
+
     } bugs = {};
 
     // state getters -- as needed.
@@ -473,12 +483,6 @@ public:
 
     void unbindEverything() noexcept;
     void synchronizeStateAndCache(size_t index) noexcept;
-    void setEs2UniformBinding(size_t index, GLuint id, void const* data, uint16_t age) noexcept {
-        mUniformBindings[index] = { id, data, age };
-    }
-    auto getEs2UniformBinding(size_t index) const noexcept {
-        return mUniformBindings[index];
-    }
 
 #ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
     GLuint getSamplerSlow(SamplerParams sp) const noexcept;
@@ -505,9 +509,6 @@ private:
     std::vector<std::function<void(OpenGLContext&)>> mDestroyWithNormalContext;
     RenderPrimitive mDefaultVAO;
     std::optional<GLuint> mDefaultFbo[2];
-    std::array<
-            std::tuple<GLuint, void const*, uint16_t>,
-            CONFIG_UNIFORM_BINDING_COUNT> mUniformBindings = {};
     mutable tsl::robin_map<SamplerParams, GLuint,
             SamplerParams::Hasher, SamplerParams::EqualTo> mSamplerMap;
 
@@ -558,6 +559,9 @@ private:
             {   bugs.delay_fbo_destruction,
                     "delay_fbo_destruction",
                     ""},
+            {   bugs.rebind_buffer_after_deletion,
+                    "rebind_buffer_after_deletion",
+                    ""},
             {   bugs.force_feature_level0,
                     "force_feature_level0",
                     ""},
@@ -627,6 +631,7 @@ constexpr size_t OpenGLContext::getIndexForCap(GLenum cap) noexcept { //NOLINT
 #ifdef BACKEND_OPENGL_VERSION_GL
         case GL_PROGRAM_POINT_SIZE:             index = 10; break;
 #endif
+        case GL_DEPTH_CLAMP:                    index = 11; break;
         default: break;
     }
     assert_invariant(index < state.enables.caps.size());

filament/backend/src/opengl/OpenGLDriver.h

@@ -21,6 +21,8 @@
 #include "OpenGLContext.h"
 #include "OpenGLTimerQuery.h"
 #include "GLBufferObject.h"
+#include "GLDescriptorSet.h"
+#include "GLDescriptorSetLayout.h"
 #include "GLTexture.h"
 #include "ShaderCompilerService.h"
 
@@ -36,6 +38,7 @@
 #include "private/backend/Driver.h"
 #include "private/backend/HandleAllocator.h"
 
+#include <utils/bitset.h>
 #include <utils/FixedCapacityVector.h>
 #include <utils/compiler.h>
 #include <utils/debug.h>
@@ -52,6 +55,7 @@
 #include <tuple>
 #include <type_traits>
 #include <utility>
+#include <variant>
 #include <vector>
 
 #include <stddef.h>
@@ -81,7 +85,7 @@ public:
             const Platform::DriverConfig& driverConfig) noexcept;
 
     class DebugMarker {
-        OpenGLDriver& driver;
+        UTILS_UNUSED OpenGLDriver& driver;
     public:
         DebugMarker(OpenGLDriver& driver, const char* string) noexcept;
         ~DebugMarker() noexcept;
@@ -123,16 +127,6 @@ public:
         } gl;
     };
 
-    struct GLSamplerGroup : public HwSamplerGroup {
-        using HwSamplerGroup::HwSamplerGroup;
-        struct Entry {
-            GLTexture const* texture = nullptr;
-            GLuint sampler = 0u;
-        };
-        utils::FixedCapacityVector<Entry> textureUnitEntries;
-        explicit GLSamplerGroup(size_t size) noexcept : textureUnitEntries(size) { }
-    };
-
     struct GLRenderPrimitive : public HwRenderPrimitive {
         using HwRenderPrimitive::HwRenderPrimitive;
         OpenGLContext::RenderPrimitive gl;
@@ -145,6 +139,10 @@ public:
 
     using GLTimerQuery = filament::backend::GLTimerQuery;
 
+    using GLDescriptorSetLayout = filament::backend::GLDescriptorSetLayout;
+
+    using GLDescriptorSet = filament::backend::GLDescriptorSet;
+
     struct GLStream : public HwStream {
         using HwStream::HwStream;
         struct Info {
@@ -256,6 +254,11 @@ private:
         return mHandleAllocator.handle_cast<Dp, B>(handle);
     }
 
+    template<typename B>
+    bool is_valid(Handle<B>& handle) {
+        return mHandleAllocator.is_valid(handle);
+    }
+
     template<typename Dp, typename B>
     inline typename std::enable_if_t<
             std::is_pointer_v<Dp> &&
@@ -312,10 +315,6 @@ private:
     void resolvePass(ResolveAction action, GLRenderTarget const* rt,
             TargetBufferFlags discardFlags) noexcept;
 
-    const std::array<GLSamplerGroup*, Program::SAMPLER_BINDING_COUNT>& getSamplerBindings() const {
-        return mSamplerBindings;
-    }
-
     using AttachmentArray = std::array<GLenum, MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT + 2>;
     static GLsizei getAttachments(AttachmentArray& attachments, TargetBufferFlags buffers,
             bool isDefaultFramebuffer) noexcept;
@@ -328,8 +327,16 @@ private:
     GLboolean mRenderPassStencilWrite{};
 
     GLRenderPrimitive const* mBoundRenderPrimitive = nullptr;
+    OpenGLProgram* mBoundProgram = nullptr;
     bool mValidProgram = false;
+    utils::bitset8 mInvalidDescriptorSetBindings;
+    utils::bitset8 mInvalidDescriptorSetBindingOffsets;
+    void updateDescriptors(utils::bitset8 invalidDescriptorSets) noexcept;
 
+    struct {
+        backend::DescriptorSetHandle dsh;
+        std::array<uint32_t, CONFIG_UNIFORM_BINDING_COUNT> offsets;
+    } mBoundDescriptorSets[MAX_DESCRIPTOR_SET_COUNT];
 
     void clearWithRasterPipe(TargetBufferFlags clearFlags,
             math::float4 const& linearColor, GLfloat depth, GLint stencil) noexcept;
@@ -341,9 +348,6 @@ private:
     // ES2 only. Uniform buffer emulation binding points
     GLuint mLastAssignedEmulatedUboId = 0;
 
-    // sampler buffer binding points (nullptr if not used)
-    std::array<GLSamplerGroup*, Program::SAMPLER_BINDING_COUNT> mSamplerBindings = {};   // 4 pointers
-
     // this must be accessed from the driver thread only
     std::vector<GLTexture*> mTexturesWithStreamsAttached;
 
@@ -354,8 +358,6 @@ private:
     void detachStream(GLTexture* t) noexcept;
     void replaceStream(GLTexture* t, GLStream* stream) noexcept;
 
-    void updateTextureLodRange(GLTexture* texture, int8_t targetLevel) noexcept;
-
 #ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
     // tasks executed on the main thread after the fence signaled
     void whenGpuCommandsComplete(const std::function<void()>& fn) noexcept;
@@ -379,6 +381,7 @@ private:
     bool mRec709OutputColorspace = false;
 
     PushConstantBundle* mCurrentPushConstants = nullptr;
+    PipelineLayout::SetLayout mCurrentSetLayout;
 };
 
 // ------------------------------------------------------------------------------------------------

filament/backend/src/opengl/OpenGLProgram.cpp

@@ -17,24 +17,30 @@
 #include "OpenGLProgram.h"
 
 #include "GLUtils.h"
+#include "GLTexture.h"
 #include "OpenGLDriver.h"
 #include "ShaderCompilerService.h"
 
+#include <backend/DriverEnums.h>
 #include <backend/Program.h>
+#include <backend/Handle.h>
 
-#include <private/backend/BackendUtils.h>
-
-#include <utils/debug.h>
+#include <utils/BitmaskEnum.h>
 #include <utils/compiler.h>
+#include <utils/debug.h>
+#include <utils/FixedCapacityVector.h>
 #include <utils/Log.h>
 #include <utils/Systrace.h>
 
+#include <algorithm>
 #include <array>
+#include <algorithm>
+#include <new>
 #include <string_view>
 #include <utility>
-#include <new>
 
 #include <stddef.h>
+#include <stdint.h>
 
 namespace filament::backend {
 
@@ -43,9 +49,8 @@ using namespace utils;
 using namespace backend;
 
 struct OpenGLProgram::LazyInitializationData {
-    Program::UniformBlockInfo uniformBlockInfo;
-    Program::SamplerGroupInfo samplerGroupInfo;
-    std::array<Program::UniformInfo, Program::UNIFORM_BINDING_COUNT> bindingUniformInfo;
+    Program::DescriptorSetInfo descriptorBindings;
+    Program::BindingUniformsInfo bindingUniformInfo;
     utils::FixedCapacityVector<Program::PushConstant> vertexPushConstants;
     utils::FixedCapacityVector<Program::PushConstant> fragmentPushConstants;
 };
@@ -54,16 +59,14 @@ struct OpenGLProgram::LazyInitializationData {
 OpenGLProgram::OpenGLProgram() noexcept = default;
 
 OpenGLProgram::OpenGLProgram(OpenGLDriver& gld, Program&& program) noexcept
-        : HwProgram(std::move(program.getName())) {
+        : HwProgram(std::move(program.getName())), mRec709Location(-1) {
     auto* const lazyInitializationData = new(std::nothrow) LazyInitializationData();
-    lazyInitializationData->samplerGroupInfo = std::move(program.getSamplerGroupInfo());
     if (UTILS_UNLIKELY(gld.getContext().isES2())) {
         lazyInitializationData->bindingUniformInfo = std::move(program.getBindingUniformInfo());
-    } else {
-        lazyInitializationData->uniformBlockInfo = std::move(program.getUniformBlockBindings());
     }
     lazyInitializationData->vertexPushConstants = std::move(program.getPushConstants(ShaderStage::VERTEX));
     lazyInitializationData->fragmentPushConstants = std::move(program.getPushConstants(ShaderStage::FRAGMENT));
+    lazyInitializationData->descriptorBindings = std::move(program.getDescriptorBindings());
 
     ShaderCompilerService& compiler = gld.getShaderCompilerService();
     mToken = compiler.createProgram(name, std::move(program));
@@ -121,36 +124,86 @@ void OpenGLProgram::initializeProgramState(OpenGLContext& context, GLuint progra
 
     SYSTRACE_CALL();
 
+    // from the pipeline layout we compute a mapping from {set, binding} to {binding}
+    // for both buffers and textures
+
+    for (auto&& entry: lazyInitializationData.descriptorBindings) {
+        std::sort(entry.begin(), entry.end(),
+                [](Program::Descriptor const& lhs, Program::Descriptor const& rhs) {
+                    return lhs.binding < rhs.binding;
+                });
+    }
+
+    GLuint tmu = 0;
+    GLuint binding = 0;
+
+    // needed for samplers
+    context.useProgram(program);
+
+    UTILS_NOUNROLL
+    for (backend::descriptor_set_t set = 0; set < MAX_DESCRIPTOR_SET_COUNT; set++) {
+        for (Program::Descriptor const& entry: lazyInitializationData.descriptorBindings[set]) {
+            switch (entry.type) {
+                case DescriptorType::UNIFORM_BUFFER:
+                case DescriptorType::SHADER_STORAGE_BUFFER: {
+                    if (!entry.name.empty()) {
 #ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
-    if (!context.isES2()) {
-        // Note: This is only needed, because the layout(binding=) syntax is not permitted in glsl
-        // (ES3.0 and GL4.1). The backend needs a way to associate a uniform block to a binding point.
-        UTILS_NOUNROLL
-        for (GLuint binding = 0, n = lazyInitializationData.uniformBlockInfo.size();
-                binding < n; binding++) {
-            auto const& name = lazyInitializationData.uniformBlockInfo[binding];
-            if (!name.empty()) {
-                GLuint const index = glGetUniformBlockIndex(program, name.c_str());
-                if (index != GL_INVALID_INDEX) {
-                    glUniformBlockBinding(program, index, binding);
-                }
-                CHECK_GL_ERROR(utils::slog.e)
-            }
-        }
-    } else
+                        if (UTILS_LIKELY(!context.isES2())) {
+                            GLuint const index = glGetUniformBlockIndex(program,
+                                    entry.name.c_str());
+                            if (index != GL_INVALID_INDEX) {
+                                // this can fail if the program doesn't use this descriptor
+                                glUniformBlockBinding(program, index, binding);
+                                mBindingMap.insert(set, entry.binding,
+                                        { binding, entry.type });
+                                ++binding;
+                            }
+                        } else
 #endif
-    {
+                        {
+                            auto pos = std::find_if(lazyInitializationData.bindingUniformInfo.begin(),
+                                    lazyInitializationData.bindingUniformInfo.end(),
+                                    [&name = entry.name](const auto& item) {
+                                return std::get<1>(item) == name;
+                            });
+                            if (pos != lazyInitializationData.bindingUniformInfo.end()) {
+                                binding = std::get<0>(*pos);
+                                mBindingMap.insert(set, entry.binding, { binding, entry.type });
+                            }
+                        }
+                    }
+                    break;
+                }
+                case DescriptorType::SAMPLER: {
+                    if (!entry.name.empty()) {
+                        GLint const loc = glGetUniformLocation(program, entry.name.c_str());
+                        if (loc >= 0) {
+                            // this can fail if the program doesn't use this descriptor
+                            mBindingMap.insert(set, entry.binding, { tmu, entry.type });
+                            glUniform1i(loc, GLint(tmu));
+                            ++tmu;
+                        }
+                    }
+                    break;
+                }
+                case DescriptorType::INPUT_ATTACHMENT:
+                    break;
+            }
+            CHECK_GL_ERROR(utils::slog.e)
+        }
+    }
+
+    if (context.isES2()) {
         // ES2 initialization of (fake) UBOs
         UniformsRecord* const uniformsRecords = new(std::nothrow) UniformsRecord[Program::UNIFORM_BINDING_COUNT];
         UTILS_NOUNROLL
-        for (GLuint binding = 0, n = Program::UNIFORM_BINDING_COUNT; binding < n; binding++) {
-            Program::UniformInfo& uniforms = lazyInitializationData.bindingUniformInfo[binding];
-            uniformsRecords[binding].locations.reserve(uniforms.size());
-            uniformsRecords[binding].locations.resize(uniforms.size());
+        for (auto&& [index, name, uniforms] : lazyInitializationData.bindingUniformInfo) {
+            uniformsRecords[index].locations.reserve(uniforms.size());
+            uniformsRecords[index].locations.resize(uniforms.size());
             for (size_t j = 0, c = uniforms.size(); j < c; j++) {
                 GLint const loc = glGetUniformLocation(program, uniforms[j].name.c_str());
-                uniformsRecords[binding].locations[j] = loc;
-                if (UTILS_UNLIKELY(binding == 0)) {
+                uniformsRecords[index].locations[j] = loc;
+                if (UTILS_UNLIKELY(index == 0)) {
                     // This is a bit of a gross hack here, we stash the location of
                     // "frameUniforms.rec709", which obviously the backend shouldn't know about,
                     // which is used for emulating the "rec709" colorspace in the shader.
@@ -162,51 +215,11 @@ void OpenGLProgram::initializeProgramState(OpenGLContext& context, GLuint progra
                     }
                 }
             }
-            uniformsRecords[binding].uniforms = std::move(uniforms);
+            uniformsRecords[index].uniforms = std::move(uniforms);
         }
         mUniformsRecords = uniformsRecords;
     }
 
-    uint8_t usedBindingCount = 0;
-    uint8_t tmu = 0;
-
-    UTILS_NOUNROLL
-    for (size_t i = 0, c = lazyInitializationData.samplerGroupInfo.size(); i < c; i++) {
-        auto const& samplers = lazyInitializationData.samplerGroupInfo[i].samplers;
-        if (samplers.empty()) {
-            // this binding point doesn't have any samplers, skip it.
-            continue;
-        }
-
-        // keep this in the loop, so we skip it in the rare case a program doesn't have
-        // sampler. The context cache will prevent repeated calls to GL.
-        context.useProgram(program);
-
-        bool atLeastOneSamplerUsed = false;
-        UTILS_NOUNROLL
-        for (const Program::Sampler& sampler: samplers) {
-            // find its location and associate a TMU to it
-            GLint const loc = glGetUniformLocation(program, sampler.name.c_str());
-            if (loc >= 0) {
-                // this can fail if the program doesn't use this sampler
-                glUniform1i(loc, tmu);
-                atLeastOneSamplerUsed = true;
-            }
-            tmu++;
-        }
-
-        // if this program doesn't use any sampler from this HwSamplerGroup, just cancel the
-        // whole group.
-        if (atLeastOneSamplerUsed) {
-            // Cache the sampler uniform locations for each interface block
-            mUsedSamplerBindingPoints[usedBindingCount] = i;
-            usedBindingCount++;
-        } else {
-            tmu -= samplers.size();
-        }
-    }
-    mUsedBindingsCount = usedBindingCount;
-
     auto& vertexConstants = lazyInitializationData.vertexPushConstants;
     auto& fragmentConstants = lazyInitializationData.fragmentPushConstants;
 
@@ -223,40 +236,8 @@ void OpenGLProgram::initializeProgramState(OpenGLContext& context, GLuint progra
     }
 }
 
-void OpenGLProgram::updateSamplers(OpenGLDriver* const gld) const noexcept {
-    using GLTexture = OpenGLDriver::GLTexture;
-
-#ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
-    bool const es2 = gld->getContext().isES2();
-#endif
-
-    // cache a few member variable locally, outside the loop
-    auto const& UTILS_RESTRICT samplerBindings = gld->getSamplerBindings();
-    auto const& UTILS_RESTRICT usedBindingPoints = mUsedSamplerBindingPoints;
-
-    for (uint8_t i = 0, tmu = 0, n = mUsedBindingsCount; i < n; i++) {
-        size_t const binding = usedBindingPoints[i];
-        assert_invariant(binding < Program::SAMPLER_BINDING_COUNT);
-        auto const * const sb = samplerBindings[binding];
-        assert_invariant(sb);
-        if (!sb) continue; // should never happen, this would be a user error.
-        for (uint8_t j = 0, m = sb->textureUnitEntries.size(); j < m; ++j, ++tmu) { // "<=" on purpose here
-            const GLTexture* const t = sb->textureUnitEntries[j].texture;
-            if (t) { // program may not use all samplers of sampler group
-                gld->bindTexture(tmu, t);
-#ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
-                if (UTILS_LIKELY(!es2)) {
-                    GLuint const s = sb->textureUnitEntries[j].sampler;
-                    gld->bindSampler(tmu, s);
-                }
-#endif
-            }
-        }
-    }
-    CHECK_GL_ERROR(utils::slog.e)
-}
-
-void OpenGLProgram::updateUniforms(uint32_t index, GLuint id, void const* buffer, uint16_t age) noexcept {
+void OpenGLProgram::updateUniforms(
+        uint32_t index, GLuint id, void const* buffer, uint16_t age) const noexcept {
     assert_invariant(mUniformsRecords);
     assert_invariant(buffer);
 

filament/backend/src/opengl/OpenGLProgram.h

@@ -19,17 +19,20 @@
 
 #include "DriverBase.h"
 
+#include "BindingMap.h"
 #include "OpenGLContext.h"
 #include "ShaderCompilerService.h"
 
 #include <private/backend/Driver.h>
+
+#include <backend/DriverEnums.h>
 #include <backend/Program.h>
 
+#include <utils/bitset.h>
 #include <utils/compiler.h>
 #include <utils/FixedCapacityVector.h>
 #include <utils/Slice.h>
 
-#include <array>
 #include <limits>
 
 #include <stddef.h>
@@ -53,37 +56,41 @@ public:
 
     bool isValid() const noexcept { return mToken || gl.program != 0; }
 
-    void use(OpenGLDriver* const gld, OpenGLContext& context) noexcept {
-        if (UTILS_UNLIKELY(!gl.program)) {
+    bool use(OpenGLDriver* const gld, OpenGLContext& context) noexcept {
+        // both non-null is impossible by construction
+        assert_invariant(!mToken || !gl.program);
+
+        if (UTILS_UNLIKELY(mToken && !gl.program)) {
+            // first time a program is used
             initialize(*gld);
         }
 
-        context.useProgram(gl.program);
-        if (UTILS_UNLIKELY(mUsedBindingsCount)) {
-            // We rely on GL state tracking to avoid unnecessary glBindTexture / glBindSampler
-            // calls.
-
-            // we need to do this if:
-            // - the content of mSamplerBindings has changed
-            // - the content of any bound sampler buffer has changed
-            // ... since last time we used this program
-
-            // Turns out the former might be relatively cheap to check, the latter requires
-            // a bit less. Compared to what updateSamplers() actually does, which is
-            // pretty little, I'm not sure if we'll get ahead.
-
-            updateSamplers(gld);
+        if (UTILS_UNLIKELY(!gl.program)) {
+            // compilation failed (token should be null)
+            assert_invariant(!mToken);
+            return false;
         }
+
+        context.useProgram(gl.program);
+        return true;
+    }
+
+    GLuint getBufferBinding(descriptor_set_t set, descriptor_binding_t binding) const noexcept {
+        return mBindingMap.get(set, binding);
+    }
+
+    GLuint getTextureUnit(descriptor_set_t set, descriptor_binding_t binding) const noexcept {
+        return mBindingMap.get(set, binding);
+    }
+
+    utils::bitset64 getActiveDescriptors(descriptor_set_t set) const noexcept {
+        return mBindingMap.getActiveDescriptors(set);
     }
 
     // For ES2 only
-    void updateUniforms(uint32_t index, GLuint id, void const* buffer, uint16_t age) noexcept;
+    void updateUniforms(uint32_t index, GLuint id, void const* buffer, uint16_t age) const noexcept;
     void setRec709ColorSpace(bool rec709) const noexcept;
 
-    struct {
-        GLuint program = 0;
-    } gl;                                               // 4 bytes
-
     PushConstantBundle getPushConstants() {
         auto fragBegin = mPushConstants.begin() + mPushConstantFragmentStageOffset;
         return {
@@ -101,22 +108,15 @@ private:
     void initializeProgramState(OpenGLContext& context, GLuint program,
             LazyInitializationData& lazyInitializationData) noexcept;
 
-    void updateSamplers(OpenGLDriver* gld) const noexcept;
-
-    // number of bindings actually used by this program
-    std::array<uint8_t, Program::SAMPLER_BINDING_COUNT> mUsedSamplerBindingPoints;   // 4 bytes
+    BindingMap mBindingMap;     // 8 bytes + out-of-line 256 bytes
 
     ShaderCompilerService::program_token_t mToken{};    // 16 bytes
 
-    uint8_t mUsedBindingsCount = 0u;                    // 1 byte
-    UTILS_UNUSED uint8_t padding[2] = {};               // 2 byte
-
-    // Push constant array offset for fragment stage constants.
-    uint8_t mPushConstantFragmentStageOffset = 0u;      // 1 byte
+    // Note that this can be replaced with a raw pointer and an uint8_t (for size) to reduce the
+    // size of the container to 9 bytes if there is a need in the future.
+    utils::FixedCapacityVector<std::pair<GLint, ConstantType>> mPushConstants;// 16 bytes
 
     // only needed for ES2
-    GLint mRec709Location = -1;                         // 4 bytes
-
     using LocationInfo = utils::FixedCapacityVector<GLint>;
     struct UniformsRecord {
         Program::UniformInfo uniforms;
@@ -124,15 +124,20 @@ private:
         mutable GLuint id = 0;
         mutable uint16_t age = std::numeric_limits<uint16_t>::max();
     };
-    UniformsRecord const* mUniformsRecords = nullptr;               // 8 bytes
+    UniformsRecord const* mUniformsRecords = nullptr;
+    GLint mRec709Location : 24;     // 4 bytes
 
-    // Note that this can be replaced with a raw pointer and an uint8_t (for size) to reduce the
-    // size of the container to 9 bytes if there is a need in the future.
-    utils::FixedCapacityVector<std::pair<GLint, ConstantType>> mPushConstants;// 16 bytes
+    // Push constant array offset for fragment stage constants.
+    GLint mPushConstantFragmentStageOffset : 8;      // 1 byte
+
+public:
+    struct {
+        GLuint program = 0;
+    } gl;                                               // 4 bytes
 };
 
-// if OpenGLProgram is larger tha 64 bytes, it'll fall in a larger Handle bucket.
-static_assert(sizeof(OpenGLProgram) <= 64); // currently 64 bytes
+// if OpenGLProgram is larger than 96 bytes, it'll fall in a larger Handle bucket.
+static_assert(sizeof(OpenGLProgram) <= 96); // currently 96 bytes
 
 } // namespace filament::backend
 

filament/backend/src/opengl/OpenGLTimerQuery.cpp

@@ -143,21 +143,25 @@ void TimerQueryNativeFactory::endTimeElapsedQuery(OpenGLDriver& driver, GLTimerQ
 
     driver.runEveryNowAndThen([&context = mContext, weak]() -> bool {
         auto state = weak.lock();
-        if (state) {
-            GLuint available = 0;
-            context.procs.getQueryObjectuiv(state->gl.query, GL_QUERY_RESULT_AVAILABLE, &available);
-            CHECK_GL_ERROR(utils::slog.e)
-            if (!available) {
-                // we need to try this one again later
-                return false;
-            }
-            GLuint64 elapsedTime = 0;
-            // we won't end-up here if we're on ES and don't have GL_EXT_disjoint_timer_query
-            context.procs.getQueryObjectui64v(state->gl.query, GL_QUERY_RESULT, &elapsedTime);
-            state->elapsed.store((int64_t)elapsedTime, std::memory_order_relaxed);
-        } else {
-            state->elapsed.store(int64_t(TimerQueryResult::ERROR), std::memory_order_relaxed);
+        if (!state) {
+            // The timer query state has been destroyed on the way, very likely due to the IBL
+            // prefilter context destruction. We still return true to get this element removed from
+            // the query list.
+            return true;
         }
+
+        GLuint available = 0;
+        context.procs.getQueryObjectuiv(state->gl.query, GL_QUERY_RESULT_AVAILABLE, &available);
+        CHECK_GL_ERROR(utils::slog.e)
+        if (!available) {
+            // we need to try this one again later
+            return false;
+        }
+        GLuint64 elapsedTime = 0;
+        // we won't end-up here if we're on ES and don't have GL_EXT_disjoint_timer_query
+        context.procs.getQueryObjectui64v(state->gl.query, GL_QUERY_RESULT, &elapsedTime);
+        state->elapsed.store((int64_t)elapsedTime, std::memory_order_relaxed);
+
         return true;
     });
 }

filament/backend/src/opengl/ShaderCompilerService.cpp

@@ -26,16 +26,28 @@
 
 #include <utils/compiler.h>
 #include <utils/CString.h>
+#include <utils/debug.h>
+#include <utils/FixedCapacityVector.h>
 #include <utils/JobSystem.h>
 #include <utils/Log.h>
+#include <utils/ostream.h>
+#include <utils/Panic.h>
 #include <utils/Systrace.h>
 
+#include <array>
 #include <cctype>
 #include <chrono>
+#include <mutex>
+#include <memory>
 #include <string>
 #include <string_view>
+#include <thread>
+#include <utility>
 #include <variant>
 
+#include <stddef.h>
+#include <stdint.h>
+
 namespace filament::backend {
 
 using namespace utils;
@@ -359,7 +371,7 @@ ShaderCompilerService::program_token_t ShaderCompilerService::createProgram(
 GLuint ShaderCompilerService::getProgram(ShaderCompilerService::program_token_t& token) {
     GLuint const program = initialize(token);
     assert_invariant(token == nullptr);
-#ifndef FILAMENT_ENABLE_MATDBG
+#if !FILAMENT_ENABLE_MATDBG
     assert_invariant(program);
 #endif
     return program;
@@ -476,6 +488,12 @@ GLuint ShaderCompilerService::initialize(program_token_t& token) noexcept {
     // check status of program linking and shader compilation, logs error and free all resources
     // in case of error.
     bool const success = checkProgramStatus(token);
+
+    // Unless we have matdbg, we panic if a program is invalid. Otherwise, we'd get a UB.
+    // The compilation error has been logged to log.e by this point.
+    FILAMENT_CHECK_POSTCONDITION(FILAMENT_ENABLE_MATDBG || success)
+            << "OpenGL program " << token->name.c_str_safe() << " failed to link or compile";
+
     if (UTILS_LIKELY(success)) {
         program = token->gl.program;
         // no need to keep the shaders around
@@ -572,24 +590,38 @@ void ShaderCompilerService::compileShaders(OpenGLContext& context,
 
             // split shader source, so we can insert the specialization constants and the packing
             // functions
-            auto const [prolog, body] = splitShaderSource({ shader_src, shader_len });
+            auto [version, prolog, body] = splitShaderSource({ shader_src, shader_len });
 
-            const std::array<const char*, 4> sources = {
-                    prolog.data(),
-                    specializationConstantString.c_str(),
-                    packingFunctions.data(),
-                    body.data()
+            // enable ESSL 3.10 if available
+            if (context.isAtLeastGLES<3, 1>()) {
+                version = "#version 310 es\n";
+            }
+
+            std::array<std::string_view, 5> sources = {
+                version,
+                prolog,
+                specializationConstantString,
+                packingFunctions,
+                { body.data(), body.size() - 1 }  // null-terminated
             };
 
-            const std::array<GLint, 4> lengths = {
-                    (GLint)prolog.length(),
-                    (GLint)specializationConstantString.length(),
-                    (GLint)packingFunctions.length(),
-                    (GLint)body.length() - 1 // null terminated
-            };
+            // Some of the sources may be zero-length. Remove them as to avoid passing lengths of
+            // zero to glShaderSource(). glShaderSource should work with lengths of zero, but some
+            // drivers instead interpret zero as a sentinel for a null-terminated string.
+            auto partitionPoint = std::stable_partition(
+                    sources.begin(), sources.end(), [](std::string_view s) { return !s.empty(); });
+            size_t count = std::distance(sources.begin(), partitionPoint);
+
+            std::array<const char*, 5> shaderStrings;
+            std::array<GLint, 5> lengths;
+            for (size_t i = 0; i < count; i++) {
+                shaderStrings[i] = sources[i].data();
+                lengths[i] = sources[i].size();
+            }
 
             GLuint const shaderId = glCreateShader(glShaderType);
-            glShaderSource(shaderId, sources.size(), sources.data(), lengths.data());
+            glShaderSource(shaderId, count, shaderStrings.data(), lengths.data());
+
             glCompileShader(shaderId);
 
 #ifndef NDEBUG
@@ -661,6 +693,7 @@ void ShaderCompilerService::process_OVR_multiview2(OpenGLContext& context,
 
 // Tragically, OpenGL 4.1 doesn't support unpackHalf2x16 (appeared in 4.2) and
 // macOS doesn't support GL_ARB_shading_language_packing
+// Also GLES3.0 didn't have the full set of packing/unpacking functions
 std::string_view ShaderCompilerService::process_ARB_shading_language_packing(OpenGLContext& context) noexcept {
     using namespace std::literals;
 #ifdef BACKEND_OPENGL_VERSION_GL
@@ -700,31 +733,102 @@ highp uint packHalf2x16(vec2 v) {
     highp uint y = fp32tou16(v.y);
     return (y << 16u) | x;
 }
+highp uint packUnorm4x8(mediump vec4 v) {
+    v = round(clamp(v, 0.0, 1.0) * 255.0);
+    highp uint a = uint(v.x);
+    highp uint b = uint(v.y) <<  8;
+    highp uint c = uint(v.z) << 16;
+    highp uint d = uint(v.w) << 24;
+    return (a|b|c|d);
+}
+highp uint packSnorm4x8(mediump vec4 v) {
+    v = round(clamp(v, -1.0, 1.0) * 127.0);
+    highp uint a = uint((int(v.x) & 0xff));
+    highp uint b = uint((int(v.y) & 0xff)) <<  8;
+    highp uint c = uint((int(v.z) & 0xff)) << 16;
+    highp uint d = uint((int(v.w) & 0xff)) << 24;
+    return (a|b|c|d);
+}
+mediump vec4 unpackUnorm4x8(highp uint v) {
+    return vec4(float((v & 0x000000ffu)      ),
+                float((v & 0x0000ff00u) >>  8),
+                float((v & 0x00ff0000u) >> 16),
+                float((v & 0xff000000u) >> 24)) / 255.0;
+}
+mediump vec4 unpackSnorm4x8(highp uint v) {
+    int a = int(((v       ) & 0xffu) << 24u) >> 24 ;
+    int b = int(((v >>  8u) & 0xffu) << 24u) >> 24 ;
+    int c = int(((v >> 16u) & 0xffu) << 24u) >> 24 ;
+    int d = int(((v >> 24u) & 0xffu) << 24u) >> 24 ;
+    return clamp(vec4(float(a), float(b), float(c), float(d)) / 127.0, -1.0, 1.0);
+}
 )"sv;
     }
 #endif // BACKEND_OPENGL_VERSION_GL
+
+#ifdef BACKEND_OPENGL_VERSION_GLES
+    if (!context.isES2() && !context.isAtLeastGLES<3, 1>()) {
+        return R"(
+
+highp uint packUnorm4x8(mediump vec4 v) {
+    v = round(clamp(v, 0.0, 1.0) * 255.0);
+    highp uint a = uint(v.x);
+    highp uint b = uint(v.y) <<  8;
+    highp uint c = uint(v.z) << 16;
+    highp uint d = uint(v.w) << 24;
+    return (a|b|c|d);
+}
+highp uint packSnorm4x8(mediump vec4 v) {
+    v = round(clamp(v, -1.0, 1.0) * 127.0);
+    highp uint a = uint((int(v.x) & 0xff));
+    highp uint b = uint((int(v.y) & 0xff)) <<  8;
+    highp uint c = uint((int(v.z) & 0xff)) << 16;
+    highp uint d = uint((int(v.w) & 0xff)) << 24;
+    return (a|b|c|d);
+}
+mediump vec4 unpackUnorm4x8(highp uint v) {
+    return vec4(float((v & 0x000000ffu)      ),
+                float((v & 0x0000ff00u) >>  8),
+                float((v & 0x00ff0000u) >> 16),
+                float((v & 0xff000000u) >> 24)) / 255.0;
+}
+mediump vec4 unpackSnorm4x8(highp uint v) {
+    int a = int(((v       ) & 0xffu) << 24u) >> 24 ;
+    int b = int(((v >>  8u) & 0xffu) << 24u) >> 24 ;
+    int c = int(((v >> 16u) & 0xffu) << 24u) >> 24 ;
+    int d = int(((v >> 24u) & 0xffu) << 24u) >> 24 ;
+    return clamp(vec4(float(a), float(b), float(c), float(d)) / 127.0, -1.0, 1.0);
+}
+)"sv;
+    }
+#endif // BACKEND_OPENGL_VERSION_GLES
     return ""sv;
 }
 
-// split shader source code in two, the first section goes from the start to the line after the
-// last #extension, and the 2nd part goes from there to the end.
-std::array<std::string_view, 2> ShaderCompilerService::splitShaderSource(std::string_view source) noexcept {
-    auto start = source.find("#version");
-    assert_invariant(start != std::string_view::npos);
+// split shader source code in three:
+// - the version line
+// - extensions
+// - everything else
+std::array<std::string_view, 3> ShaderCompilerService::splitShaderSource(std::string_view source) noexcept {
+    auto version_start = source.find("#version");
+    assert_invariant(version_start != std::string_view::npos);
 
-    auto pos = source.rfind("\n#extension");
-    if (pos == std::string_view::npos) {
-        pos = start;
+    auto version_eol = source.find('\n', version_start) + 1;
+    assert_invariant(version_eol != std::string_view::npos);
+
+    auto prolog_start = version_eol;
+    auto prolog_eol = source.rfind("\n#extension"); // last #extension line
+    if (prolog_eol == std::string_view::npos) {
+        prolog_eol = prolog_start;
     } else {
-        ++pos;
+        prolog_eol = source.find('\n', prolog_eol + 1) + 1;
     }
+    auto body_start = prolog_eol;
 
-    auto eol = source.find('\n', pos) + 1;
-    assert_invariant(eol != std::string_view::npos);
-
-    std::string_view const version = source.substr(start, eol - start);
-    std::string_view const body = source.substr(version.length(), source.length() - version.length());
-    return { version, body };
+    std::string_view const version = source.substr(version_start, version_eol - version_start);
+    std::string_view const prolog = source.substr(prolog_start, prolog_eol - prolog_start);
+    std::string_view const body = source.substr(body_start, source.length() - body_start);
+    return { version, prolog, body };
 }
 
 /*

filament/backend/src/opengl/ShaderCompilerService.h

@@ -146,7 +146,7 @@ private:
 
     static std::string_view process_ARB_shading_language_packing(OpenGLContext& context) noexcept;
 
-    static std::array<std::string_view, 2> splitShaderSource(std::string_view source) noexcept;
+    static std::array<std::string_view, 3> splitShaderSource(std::string_view source) noexcept;
 
     static GLuint linkProgram(OpenGLContext& context,
             std::array<GLuint, Program::SHADER_TYPE_COUNT> shaders,

filament/backend/src/opengl/gl_headers.h

@@ -201,6 +201,12 @@ using namespace glext;
 #   define GL_CLIP_DISTANCE1                        0x3001
 #endif
 
+#if defined(GL_EXT_depth_clamp)
+#   define GL_DEPTH_CLAMP                           GL_DEPTH_CLAMP_EXT
+#else
+#   define GL_DEPTH_CLAMP                           0x864F
+#endif
+
 #if defined(GL_KHR_debug)
 #   define GL_DEBUG_OUTPUT                          GL_DEBUG_OUTPUT_KHR
 #   define GL_DEBUG_OUTPUT_SYNCHRONOUS              GL_DEBUG_OUTPUT_SYNCHRONOUS_KHR

filament/backend/src/opengl/platforms/PlatformEGL.cpp

@@ -155,10 +155,6 @@ Driver* PlatformEGL::createDriver(void* sharedContext, const Platform::DriverCon
     ext.egl.KHR_no_config_context = extensions.has("EGL_KHR_no_config_context");
     ext.egl.KHR_surfaceless_context = extensions.has("EGL_KHR_surfaceless_context");
     ext.egl.EXT_protected_content = extensions.has("EGL_EXT_protected_content");
-    if (ext.egl.KHR_create_context) {
-        // KHR_create_context implies KHR_surfaceless_context for ES3.x contexts
-        ext.egl.KHR_surfaceless_context = true;
-    }
 
     eglCreateSyncKHR = (PFNEGLCREATESYNCKHRPROC) eglGetProcAddress("eglCreateSyncKHR");
     eglDestroySyncKHR = (PFNEGLDESTROYSYNCKHRPROC) eglGetProcAddress("eglDestroySyncKHR");
@@ -557,6 +553,8 @@ void PlatformEGL::destroySwapChain(Platform::SwapChain* swapChain) noexcept {
     if (swapChain) {
         SwapChainEGL const* const sc = static_cast<SwapChainEGL const*>(swapChain);
         if (sc->sur != EGL_NO_SURFACE) {
+            // - if EGL_KHR_surfaceless_context is supported, mEGLDummySurface is EGL_NO_SURFACE.
+            // - this is actually a bit too aggressive, but it is a rare operation.
             egl.makeCurrent(mEGLDummySurface, mEGLDummySurface);
             eglDestroySurface(mEGLDisplay, sc->sur);
             delete sc;

filament/backend/src/opengl/platforms/PlatformEGLAndroid.cpp

@@ -19,16 +19,20 @@
 #include <backend/platforms/PlatformEGL.h>
 #include <backend/platforms/PlatformEGLAndroid.h>
 
+#include <private/backend/VirtualMachineEnv.h>
+
 #include "opengl/GLUtils.h"
 #include "ExternalStreamManagerAndroid.h"
 
 #include <android/api-level.h>
+#include <android/native_window.h>
 #include <android/hardware_buffer.h>
 
 #include <utils/android/PerformanceHintManager.h>
 
 #include <utils/compiler.h>
 #include <utils/ostream.h>
+#include <utils/Panic.h>
 #include <utils/Log.h>
 
 #include <EGL/egl.h>
@@ -40,7 +44,9 @@
 
 #include <chrono>
 #include <new>
+#include <string_view>
 
+#include <dlfcn.h>
 #include <unistd.h>
 
 #include <stddef.h>
@@ -78,45 +84,92 @@ UTILS_PRIVATE PFNEGLGETFRAMETIMESTAMPSANDROIDPROC eglGetFrameTimestampsANDROID =
 }
 using namespace glext;
 
-using EGLStream = Platform::Stream;
+// ---------------------------------------------------------------------------------------------
+
+PlatformEGLAndroid::InitializeJvmForPerformanceManagerIfNeeded::InitializeJvmForPerformanceManagerIfNeeded() {
+    // PerformanceHintManager() needs the calling thread to be a Java thread; so we need
+    // to attach this thread to the JVM before we initialize PerformanceHintManager.
+    // This should be done in PerformanceHintManager(), but libutils doesn't have access to
+    // VirtualMachineEnv.
+    if (PerformanceHintManager::isSupported()) {
+        (void)VirtualMachineEnv::get().getEnvironment();
+    }
+}
 
 // ---------------------------------------------------------------------------------------------
 
 PlatformEGLAndroid::PlatformEGLAndroid() noexcept
         : PlatformEGL(),
-          mExternalStreamManager(ExternalStreamManagerAndroid::create()) {
-
-    char scratch[PROP_VALUE_MAX + 1];
-    int length = __system_property_get("ro.build.version.release", scratch);
-    int const androidVersion = length >= 0 ? atoi(scratch) : 1;
-    if (!androidVersion) {
-        mOSVersion = 1000; // if androidVersion is 0, it means "future"
-    } else {
-        length = __system_property_get("ro.build.version.sdk", scratch);
-        mOSVersion = length >= 0 ? atoi(scratch) : 1;
+          mExternalStreamManager(ExternalStreamManagerAndroid::create()),
+          mInitializeJvmForPerformanceManagerIfNeeded(),
+          mPerformanceHintManager() {
+    mOSVersion = android_get_device_api_level();
+    if (mOSVersion < 0) {
+        mOSVersion = __ANDROID_API_FUTURE__;
     }
 
-    // This disables an ANGLE optimization on ARM, which turns out to be more costly for us
-    // see b/229017581
-    // We need to do this before we create the GL context.
-    // An alternative solution is use a system property:
-    //            __system_property_set(
-    //                    "debug.angle.feature_overrides_disabled",
-    //                    "preferSubmitAtFBOBoundary");
-    // but that would outlive this process, so the environment variable is better.
-    // We also make sure to not update the variable if it already exists.
-    // There is no harm setting this if we're not on ANGLE or ARM.
-    setenv("ANGLE_FEATURE_OVERRIDES_DISABLED", "preferSubmitAtFBOBoundary", false);
+    mNativeWindowLib = dlopen("libnativewindow.so", RTLD_LOCAL | RTLD_NOW);
+    if (mNativeWindowLib) {
+        ANativeWindow_getBuffersDefaultDataSpace =
+                (int32_t(*)(ANativeWindow*))dlsym(mNativeWindowLib,
+                        "ANativeWindow_getBuffersDefaultDataSpace");
+    }
 }
 
-PlatformEGLAndroid::~PlatformEGLAndroid() noexcept = default;
-
+PlatformEGLAndroid::~PlatformEGLAndroid() noexcept {
+    if (mNativeWindowLib) {
+        dlclose(mNativeWindowLib);
+    }
+}
 
 void PlatformEGLAndroid::terminate() noexcept {
     ExternalStreamManagerAndroid::destroy(&mExternalStreamManager);
     PlatformEGL::terminate();
 }
 
+static constexpr const std::string_view kNativeWindowInvalidMsg =
+        "ANativeWindow is invalid. It probably has been destroyed. EGL surface = ";
+
+bool PlatformEGLAndroid::makeCurrent(ContextType type,
+        SwapChain* drawSwapChain,
+        SwapChain* readSwapChain) noexcept {
+
+    // fast & safe path
+    if (UTILS_LIKELY(!mAssertNativeWindowIsValid)) {
+        return PlatformEGL::makeCurrent(type, drawSwapChain, readSwapChain);
+    }
+
+    SwapChainEGL const* const dsc = static_cast<SwapChainEGL const*>(drawSwapChain);
+    if (ANativeWindow_getBuffersDefaultDataSpace) {
+        // anw can be nullptr if we're using a pbuffer surface
+        if (UTILS_LIKELY(dsc->nativeWindow)) {
+            // this a proxy of is_valid()
+            auto result = ANativeWindow_getBuffersDefaultDataSpace(dsc->nativeWindow);
+            FILAMENT_CHECK_POSTCONDITION(result >= 0) << kNativeWindowInvalidMsg << dsc->sur;
+        }
+    } else {
+        // If we don't have ANativeWindow_getBuffersDefaultDataSpace, we revert to using the
+        // private query() call.
+        // Shadow version if the real ANativeWindow, so we can access the query() hook. Query
+        // has existed since forever, probably Android 1.0.
+        struct NativeWindow {
+            // is valid query enum value
+            enum { IS_VALID = 17 };
+            uint64_t pad[18];
+            int (* query)(ANativeWindow const*, int, int*);
+        } const* pWindow = reinterpret_cast<NativeWindow const*>(dsc->nativeWindow);
+        int isValid = 0;
+        if (UTILS_LIKELY(pWindow->query)) { // just in case it's nullptr
+            int const err = pWindow->query(dsc->nativeWindow, NativeWindow::IS_VALID, &isValid);
+            if (UTILS_LIKELY(err >= 0)) { // in case the IS_VALID enum is not recognized
+                // query call succeeded
+                FILAMENT_CHECK_POSTCONDITION(isValid) << kNativeWindowInvalidMsg << dsc->sur;
+            }
+        }
+    }
+    return PlatformEGL::makeCurrent(type, drawSwapChain, readSwapChain);
+}
+
 void PlatformEGLAndroid::beginFrame(
         int64_t monotonic_clock_ns,
         int64_t refreshIntervalNs,
@@ -173,6 +226,8 @@ Driver* PlatformEGLAndroid::createDriver(void* sharedContext,
                 "eglGetFrameTimestampsANDROID");
     }
 
+    mAssertNativeWindowIsValid = driverConfig.assertNativeWindowIsValid;
+
     return driver;
 }
 

filament/backend/src/opengl/platforms/PlatformOSMesa.cpp

@@ -0,0 +1,149 @@
+/*
+ * Copyright (C) 2024 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <backend/platforms/PlatformOSMesa.h>
+
+#include <utils/Log.h>
+#include <utils/Panic.h>
+
+#include <dlfcn.h>
+#include <memory>
+
+namespace filament::backend {
+
+using namespace backend;
+
+namespace {
+
+using BackingType = GLfloat;
+#define BACKING_GL_TYPE GL_FLOAT
+
+struct OSMesaSwapchain {
+    OSMesaSwapchain(uint32_t width, uint32_t height)
+        : width(width),
+          height(height),
+          buffer(new uint8_t[width * height * 4 * sizeof(BackingType)]) {}
+
+    uint32_t width = 0;
+    uint32_t height = 0;
+    std::unique_ptr<uint8_t[]> buffer;
+};
+
+struct OSMesaAPI {
+private:
+    using CreateContextFunc = OSMesaContext (*)(GLenum format, OSMesaContext);
+    using DestroyContextFunc = GLboolean (*)(OSMesaContext);
+    using MakeCurrentFunc = GLboolean (*)(OSMesaContext ctx, void* buffer, GLenum type,
+            GLsizei width, GLsizei height);
+    using GetProcAddressFunc = OSMESAproc (*)(const char* funcName);
+
+public:
+    CreateContextFunc OSMesaCreateContext;
+    DestroyContextFunc OSMesaDestroyContext;
+    MakeCurrentFunc OSMesaMakeCurrent;
+    GetProcAddressFunc OSMesaGetProcAddress;
+
+    OSMesaAPI() {
+        constexpr char const* libraryNames[] = {"libOSMesa.so", "libosmesa.so"};
+        for (char const* libName: libraryNames) {
+            mLib = dlopen(libName, RTLD_GLOBAL | RTLD_NOW);
+            if (mLib) {
+                break;
+            }
+        }
+        FILAMENT_CHECK_PRECONDITION(mLib)
+                << "Unable to dlopen libOSMesa to create a software GL context";
+
+        OSMesaGetProcAddress = (GetProcAddressFunc) dlsym(mLib, "OSMesaGetProcAddress");
+
+        OSMesaCreateContext = (CreateContextFunc) OSMesaGetProcAddress("OSMesaCreateContext");
+        OSMesaDestroyContext =
+                (DestroyContextFunc) OSMesaGetProcAddress("OSMesaDestroyContext");
+        OSMesaMakeCurrent = (MakeCurrentFunc) OSMesaGetProcAddress("OSMesaMakeCurrent");
+    }
+
+    ~OSMesaAPI() {
+        dlclose(mLib);
+    }
+private:
+    void* mLib = nullptr;
+};
+
+}// anonymous namespace
+
+Driver* PlatformOSMesa::createDriver(void* const sharedGLContext,
+        const DriverConfig& driverConfig) noexcept {
+    OSMesaAPI* api = new OSMesaAPI();
+    mOsMesaApi = api;
+
+    FILAMENT_CHECK_PRECONDITION(sharedGLContext == nullptr)
+            << "shared GL context is not supported with PlatformOSMesa";
+    mContext = api->OSMesaCreateContext(GL_RGBA, NULL);
+
+    // We need to do a no-op makecurrent here so that the context will be in a correct state before
+    // any GL calls.
+    auto chain = createSwapChain(1, 1, 0);
+    makeCurrent(ContextType::UNPROTECTED, chain, nullptr);
+    destroySwapChain(chain);
+
+    int result = bluegl::bind();
+    FILAMENT_CHECK_POSTCONDITION(!result) << "Unable to load OpenGL entry points.";
+
+    return OpenGLPlatform::createDefaultDriver(this, sharedGLContext, driverConfig);
+}
+
+void PlatformOSMesa::terminate() noexcept {
+    OSMesaAPI* api = (OSMesaAPI*) mOsMesaApi;
+    api->OSMesaDestroyContext(mContext);
+    delete api;
+    mOsMesaApi = nullptr;
+
+    bluegl::unbind();
+}
+
+Platform::SwapChain* PlatformOSMesa::createSwapChain(void* nativeWindow, uint64_t flags) noexcept {
+    FILAMENT_CHECK_POSTCONDITION(false) << "Cannot create non-headless swapchain";
+    return (SwapChain*) nativeWindow;
+}
+
+Platform::SwapChain* PlatformOSMesa::createSwapChain(uint32_t width, uint32_t height,
+        uint64_t flags) noexcept {
+    OSMesaSwapchain* swapchain = new OSMesaSwapchain(width, height);
+    return (SwapChain*) swapchain;
+}
+
+void PlatformOSMesa::destroySwapChain(Platform::SwapChain* swapChain) noexcept {
+    OSMesaSwapchain* impl = (OSMesaSwapchain*) swapChain;
+    delete impl;
+}
+
+bool PlatformOSMesa::makeCurrent(ContextType type, SwapChain* drawSwapChain,
+        SwapChain* readSwapChain) noexcept {
+    OSMesaAPI* api = (OSMesaAPI*) mOsMesaApi;
+    OSMesaSwapchain* impl = (OSMesaSwapchain*) drawSwapChain;
+
+    auto result = api->OSMesaMakeCurrent(mContext, (BackingType*) impl->buffer.get(),
+            BACKING_GL_TYPE, impl->width, impl->height);
+    FILAMENT_CHECK_POSTCONDITION(result == GL_TRUE) << "OSMesaMakeCurrent failed!";
+
+    return true;
+}
+
+void PlatformOSMesa::commit(Platform::SwapChain* swapChain) noexcept {
+    // No-op since we are not scanning out to a display.
+}
+
+} // namespace filament::backend

filament/backend/src/ostream.cpp

@@ -410,7 +410,6 @@ io::ostream& operator<<(io::ostream& out, const RasterState& rs) {
 io::ostream& operator<<(io::ostream& out, const TargetBufferInfo& tbi) {
     return out << "TargetBufferInfo{"
     << "handle=" << tbi.handle
-    << ", baseViewIndex=" << tbi.baseViewIndex
     << ", level=" << tbi.level
     << ", layer=" << tbi.layer << "}";
 }

filament/backend/src/vulkan/VulkanBlitter.cpp

@@ -15,6 +15,7 @@
  */
 
 #include "VulkanBlitter.h"
+#include "VulkanCommands.h"
 #include "VulkanContext.h"
 #include "VulkanFboCache.h"
 #include "VulkanHandles.h"
@@ -33,11 +34,12 @@ namespace filament::backend {
 
 namespace {
 
-inline void blitFast(const VkCommandBuffer cmdbuffer, VkImageAspectFlags aspect, VkFilter filter,
+inline void blitFast(VulkanCommandBuffer* commands, VkImageAspectFlags aspect, VkFilter filter,
         VulkanAttachment src, VulkanAttachment dst,
         const VkOffset3D srcRect[2], const VkOffset3D dstRect[2]) {
+    VkCommandBuffer const cmdbuf = commands->buffer();
     if constexpr (FVK_ENABLED(FVK_DEBUG_BLITTER)) {
-        utils::slog.d << "Fast blit from=" << src.texture->getVkImage() << ",level=" << (int) src.level
+        FVK_LOGD << "Fast blit from=" << src.texture->getVkImage() << ",level=" << (int) src.level
                       << " layout=" << src.getLayout()
                       << " to=" << dst.texture->getVkImage() << ",level=" << (int) dst.level
                       << " layout=" << dst.getLayout() << utils::io::endl;
@@ -49,8 +51,8 @@ inline void blitFast(const VkCommandBuffer cmdbuffer, VkImageAspectFlags aspect,
     VulkanLayout oldSrcLayout = src.getLayout();
     VulkanLayout oldDstLayout = dst.getLayout();
 
-    src.texture->transitionLayout(cmdbuffer, srcRange, VulkanLayout::TRANSFER_SRC);
-    dst.texture->transitionLayout(cmdbuffer, dstRange, VulkanLayout::TRANSFER_DST);
+    src.texture->transitionLayout(commands, srcRange, VulkanLayout::TRANSFER_SRC);
+    dst.texture->transitionLayout(commands, dstRange, VulkanLayout::TRANSFER_DST);
 
     const VkImageBlit blitRegions[1] = {{
             .srcSubresource = { aspect, src.level, src.layer, 1 },
@@ -58,25 +60,26 @@ inline void blitFast(const VkCommandBuffer cmdbuffer, VkImageAspectFlags aspect,
             .dstSubresource = { aspect, dst.level, dst.layer, 1 },
             .dstOffsets = { dstRect[0], dstRect[1] },
     }};
-    vkCmdBlitImage(cmdbuffer,
+    vkCmdBlitImage(cmdbuf,
             src.getImage(), imgutil::getVkLayout(VulkanLayout::TRANSFER_SRC),
             dst.getImage(), imgutil::getVkLayout(VulkanLayout::TRANSFER_DST),
             1, blitRegions, filter);
 
     if (oldSrcLayout == VulkanLayout::UNDEFINED) {
-        oldSrcLayout = imgutil::getDefaultLayout(src.texture->usage);
+        oldSrcLayout = src.texture->getDefaultLayout();
     }
     if (oldDstLayout == VulkanLayout::UNDEFINED) {
-        oldDstLayout = imgutil::getDefaultLayout(dst.texture->usage);
+        oldDstLayout = dst.texture->getDefaultLayout();
     }
-    src.texture->transitionLayout(cmdbuffer, srcRange, oldSrcLayout);
-    dst.texture->transitionLayout(cmdbuffer, dstRange, oldDstLayout);
+    src.texture->transitionLayout(commands, srcRange, oldSrcLayout);
+    dst.texture->transitionLayout(commands, dstRange, oldDstLayout);
 }
 
-inline void resolveFast(const VkCommandBuffer cmdbuffer, VkImageAspectFlags aspect,
+inline void resolveFast(VulkanCommandBuffer* commands, VkImageAspectFlags aspect,
         VulkanAttachment src, VulkanAttachment dst) {
+    VkCommandBuffer const cmdbuffer = commands->buffer();
     if constexpr (FVK_ENABLED(FVK_DEBUG_BLITTER)) {
-        utils::slog.d << "Fast blit from=" << src.texture->getVkImage() << ",level=" << (int) src.level
+        FVK_LOGD << "Fast blit from=" << src.texture->getVkImage() << ",level=" << (int) src.level
                       << " layout=" << src.getLayout()
                       << " to=" << dst.texture->getVkImage() << ",level=" << (int) dst.level
                       << " layout=" << dst.getLayout() << utils::io::endl;
@@ -88,8 +91,8 @@ inline void resolveFast(const VkCommandBuffer cmdbuffer, VkImageAspectFlags aspe
     VulkanLayout oldSrcLayout = src.getLayout();
     VulkanLayout oldDstLayout = dst.getLayout();
 
-    src.texture->transitionLayout(cmdbuffer, srcRange, VulkanLayout::TRANSFER_SRC);
-    dst.texture->transitionLayout(cmdbuffer, dstRange, VulkanLayout::TRANSFER_DST);
+    src.texture->transitionLayout(commands, srcRange, VulkanLayout::TRANSFER_SRC);
+    dst.texture->transitionLayout(commands, dstRange, VulkanLayout::TRANSFER_DST);
 
     assert_invariant(
             aspect != VK_IMAGE_ASPECT_DEPTH_BIT && "Resolve with depth is not yet supported.");
@@ -106,13 +109,13 @@ inline void resolveFast(const VkCommandBuffer cmdbuffer, VkImageAspectFlags aspe
             1, resolveRegions);
 
     if (oldSrcLayout == VulkanLayout::UNDEFINED) {
-        oldSrcLayout = imgutil::getDefaultLayout(src.texture->usage);
+        oldSrcLayout = src.texture->getDefaultLayout();
     }
     if (oldDstLayout == VulkanLayout::UNDEFINED) {
-        oldDstLayout = imgutil::getDefaultLayout(dst.texture->usage);
+        oldDstLayout = dst.texture->getDefaultLayout();
     }
-    src.texture->transitionLayout(cmdbuffer, srcRange, oldSrcLayout);
-    dst.texture->transitionLayout(cmdbuffer, dstRange, oldDstLayout);
+    src.texture->transitionLayout(commands, srcRange, oldSrcLayout);
+    dst.texture->transitionLayout(commands, dstRange, oldDstLayout);
 }
 
 struct BlitterUniforms {
@@ -149,10 +152,9 @@ void VulkanBlitter::resolve(VulkanAttachment dst, VulkanAttachment src) {
 #endif
 
     VulkanCommandBuffer& commands = mCommands->get();
-    VkCommandBuffer const cmdbuffer = commands.buffer();
     commands.acquire(src.texture);
     commands.acquire(dst.texture);
-    resolveFast(cmdbuffer, aspect, src, dst);
+    resolveFast(&commands, aspect, src, dst);
 }
 
 void VulkanBlitter::blit(VkFilter filter,
@@ -175,10 +177,9 @@ void VulkanBlitter::blit(VkFilter filter,
     // src and dst should have the same aspect here
     VkImageAspectFlags const aspect = src.texture->getImageAspect();
     VulkanCommandBuffer& commands = mCommands->get();
-    VkCommandBuffer const cmdbuffer = commands.buffer();
     commands.acquire(src.texture);
     commands.acquire(dst.texture);
-    blitFast(cmdbuffer, aspect, filter, src, dst, srcRectPair, dstRectPair);
+    blitFast(&commands, aspect, filter, src, dst, srcRectPair, dstRectPair);
 }
 
 void VulkanBlitter::terminate() noexcept {

filament/backend/src/vulkan/VulkanBuffer.cpp

@@ -28,6 +28,7 @@ VulkanBuffer::VulkanBuffer(VmaAllocator allocator, VulkanStagePool& stagePool,
     : mAllocator(allocator),
       mStagePool(stagePool),
       mUsage(usage),
+	  mUpdatedOffset(0),
       mUpdatedBytes(0) {
     // for now make sure that only 1 bit is set in usage
     // (because loadFromCpu() assumes that somewhat)
@@ -80,6 +81,7 @@ void VulkanBuffer::loadFromCpu(VkCommandBuffer cmdbuf, const void* cpuData, uint
             .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
             .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
             .buffer = mGpuBuffer,
+			.offset = mUpdatedOffset,
             .size = mUpdatedBytes,
         };
         vkCmdPipelineBarrier(cmdbuf, srcStage, VK_PIPELINE_STAGE_TRANSFER_BIT, 0, 0, nullptr, 1,
@@ -93,6 +95,7 @@ void VulkanBuffer::loadFromCpu(VkCommandBuffer cmdbuf, const void* cpuData, uint
     };
     vkCmdCopyBuffer(cmdbuf, stage->buffer, mGpuBuffer, 1, &region);
 
+	mUpdatedOffset = byteOffset;
     mUpdatedBytes = numBytes;
 
     // Firstly, ensure that the copy finishes before the next draw call.

filament/backend/src/vulkan/VulkanBuffer.h

@@ -42,6 +42,7 @@ private:
     VmaAllocation mGpuMemory = VK_NULL_HANDLE;
     VkBuffer mGpuBuffer = VK_NULL_HANDLE;
     VkBufferUsageFlags mUsage = {};
+	uint32_t mUpdatedOffset = 0;
     uint32_t mUpdatedBytes = 0;
 };
 

filament/backend/src/vulkan/VulkanCommands.cpp

@@ -178,7 +178,7 @@ VulkanCommandBuffer& VulkanCommands::get() {
     // presenting the swap chain or waiting on a fence.
     while (mAvailableBufferCount == 0) {
 #if FVK_ENABLED(FVK_DEBUG_COMMAND_BUFFER)
-        slog.i << "VulkanCommands has stalled. "
+        FVK_LOGI << "VulkanCommands has stalled. "
                << "If this occurs frequently, consider increasing VK_MAX_COMMAND_BUFFERS."
                << io::endl;
 #endif
@@ -289,7 +289,7 @@ bool VulkanCommands::flush() {
     };
 
 #if FVK_ENABLED(FVK_DEBUG_COMMAND_BUFFER)
-    slog.i << "Submitting cmdbuffer=" << cmdbuffer
+    FVK_LOGI << "Submitting cmdbuffer=" << cmdbuffer
            << " wait=(" << signals[0] << ", " << signals[1] << ") "
            << " signal=" << renderingFinished
            << " fence=" << currentbuf->fence->fence
@@ -297,15 +297,15 @@ bool VulkanCommands::flush() {
 #endif
 
     auto& cmdfence = currentbuf->fence;
-    std::unique_lock<utils::Mutex> lock(cmdfence->mutex);
-    cmdfence->status.store(VK_NOT_READY);
-    UTILS_UNUSED_IN_RELEASE VkResult result = vkQueueSubmit(mQueue, 1, &submitInfo, cmdfence->fence);
-    cmdfence->condition.notify_all();
-    lock.unlock();
+    UTILS_UNUSED_IN_RELEASE VkResult result = VK_SUCCESS;
+    {
+        auto scope = cmdfence->setValue(VK_NOT_READY);
+        result = vkQueueSubmit(mQueue, 1, &submitInfo, cmdfence->getFence());
+    }
 
 #if FVK_ENABLED(FVK_DEBUG_COMMAND_BUFFER)
     if (result != VK_SUCCESS) {
-        utils::slog.d << "Failed command buffer submission result: " << result << utils::io::endl;
+        FVK_LOGD << "Failed command buffer submission result: " << result << utils::io::endl;
     }
 #endif
     assert_invariant(result == VK_SUCCESS);
@@ -320,7 +320,7 @@ VkSemaphore VulkanCommands::acquireFinishedSignal() {
     VkSemaphore semaphore = mSubmissionSignal;
     mSubmissionSignal = VK_NULL_HANDLE;
 #if FVK_ENABLED(FVK_DEBUG_COMMAND_BUFFER)
-    slog.i << "Acquiring " << semaphore << " (e.g. for vkQueuePresentKHR)" << io::endl;
+    FVK_LOGI << "Acquiring " << semaphore << " (e.g. for vkQueuePresentKHR)" << io::endl;
 #endif
     return semaphore;
 }
@@ -329,7 +329,7 @@ void VulkanCommands::injectDependency(VkSemaphore next) {
     assert_invariant(mInjectedSignal == VK_NULL_HANDLE);
     mInjectedSignal = next;
 #if FVK_ENABLED(FVK_DEBUG_COMMAND_BUFFER)
-    slog.i << "Injecting " << next << " (e.g. due to vkAcquireNextImageKHR)" << io::endl;
+    FVK_LOGI << "Injecting " << next << " (e.g. due to vkAcquireNextImageKHR)" << io::endl;
 #endif
 }
 
@@ -340,7 +340,7 @@ void VulkanCommands::wait() {
         auto wrapper = mStorage[i].get();
         if (wrapper->buffer() != VK_NULL_HANDLE
                 && mCurrentCommandBufferIndex != static_cast<int8_t>(i)) {
-            fences[count++] = wrapper->fence->fence;
+            fences[count++] = wrapper->fence->getFence();
         }
     }
     if (count > 0) {
@@ -361,12 +361,13 @@ void VulkanCommands::gc() {
         if (wrapper->buffer() == VK_NULL_HANDLE) {
             continue;
         }
-        VkResult const result = vkGetFenceStatus(mDevice, wrapper->fence->fence);
+        auto const vkfence = wrapper->fence->getFence();
+        VkResult const result = vkGetFenceStatus(mDevice, vkfence);
         if (result != VK_SUCCESS) {
             continue;
         }
-        fences[count++] = wrapper->fence->fence;
-        wrapper->fence->status.store(VK_SUCCESS);
+        fences[count++] = vkfence;
+        wrapper->fence->setValue(VK_SUCCESS);
         wrapper->reset();
         mAvailableBufferCount++;
     }
@@ -383,9 +384,9 @@ void VulkanCommands::updateFences() {
         if (wrapper->buffer() != VK_NULL_HANDLE) {
             VulkanCmdFence* fence = wrapper->fence.get();
             if (fence) {
-                VkResult status = vkGetFenceStatus(mDevice, fence->fence);
+                VkResult status = vkGetFenceStatus(mDevice, fence->getFence());
                 // This is either VK_SUCCESS, VK_NOT_READY, or VK_ERROR_DEVICE_LOST.
-                fence->status.store(status, std::memory_order_relaxed);
+                fence->setValue(status);
             }
         }
     }
@@ -398,7 +399,7 @@ void VulkanCommands::pushGroupMarker(char const* str, VulkanGroupMarkers::Timest
     // If the timestamp is not 0, then we are carrying over a marker across buffer submits.
     // If it is 0, then this is a normal marker push and we should just print debug line as usual.
     if (timestamp.time_since_epoch().count() == 0.0) {
-        utils::slog.d << "----> " << str << utils::io::endl;
+        FVK_LOGD << "----> " << str << utils::io::endl;
     }
 #endif
 
@@ -436,7 +437,7 @@ void VulkanCommands::popGroupMarker() {
         auto const [marker, startTime] = mGroupMarkers->pop();
         auto const endTime = std::chrono::high_resolution_clock::now();
         std::chrono::duration<double> diff = endTime - startTime;
-        utils::slog.d << "<---- " << marker << " elapsed: " << (diff.count() * 1000) << " ms"
+        FVK_LOGD << "<---- " << marker << " elapsed: " << (diff.count() * 1000) << " ms"
                       << utils::io::endl;
 #else
         mGroupMarkers->pop();